Move from Init() to SetUpModel() when class state is filled either with options or with XML file

Add a check if Keras version is >=2.3 If it is not and Tensorflow 2 is used then switch automatically to use Tensorflow.Keras

The support is provided by adding an option in MethodPyKeras, tf.keras=1

* update deprecated function call name to backward compatible one

* adapt convolution forward to cuDNN 8

* adapt convolution backward to cuDNN 8

* fix typo and re-declarations

* implement workspace limits, fix an algoruthm preference bug and rewrite relevant sections

* implement correct logic behind cudnn logarithm preference

* use decltype instead of auto, fix typos

* assign backward filter algo to correct place

* make it compile and support C++11

* compiles completely

Had to make sure that the GSL_CBLAS_LIBRARY variable is set in the
same way in which FindGSL.cmake would set it, and that TMVA would
explicitly wait for the completion of the GSL build.

* Add deprecation message for MethodDNN

* Improve handling of parsing of inputshape layout.
If size of input shape given is smaller than expected pad it with 1 values.
In this case different inputShape con be supported easly in case of RNN and DNN

Remove from the tutorial the usage of inputBatchLayout that is not needed anymore
Remove also using InputShapeLayout in case of dense layer networks from the tutorial

* Set correct options for Deep Learning method in TMVARegression.C

* remove an uneeded  printout statement

PyROOT/PyMVA related

These changes address the case in which Python 3 has only Interpreter
while Python 2 has both Interpreter and Development. In this case, the
Python used to build ROOT (which requires only Interpreter) will be 3,
but PyROOT (which requires also Development) will be built only for 2.
To achieve this, ROOT and PyROOT/PyMVA use now two different sets of
variables.

Documentation for the whole machinery is also added.

It is also worth pointing out that the entire machinery could be much
simplified by just requiring the following as a prerequisite for ROOT:
        - CMake >= 3.12.4
        - Python-Development package

Fix compilation of cuda with C++14 when normal ROOT is compile with C++17 which has std::string_view (#5598)

Fix it by modifying the pre-processor macros defined in RCOnfigure.h when compiling  Cuda.
A better fix would be to remove the TString dependency in the Cuda compiled  code. TString is used when doing I/O of the DeepNet layers to XML. In principle this code could be moved out of Cuda

* Fix the multi-class classification for MethodDL by using as default for the outputlayer a width equal to the number of classes.
Update multiclass tutorial to use MethodDL

* Make sure input file is generated with 2000 events. Use a different name to not mix with other tutorials

* Update input file for TMVAMulticlassApplication.C

* Update multiclass tutorials to use input file from root.cern.ch
as suggested by review comment of Stefan

Add also some other small improvements

- clean up also test CMake lists file

  Use still TCpuMatrix class (TCpu architecture) but use TMatrix for matrix multiplication instead of BLAS\
  Only dense layers architectures are supported in this case

  - Fix tutorials for the no-imt case

-   Use Higgs data file from root.cern.ch

* Update gtest to latest version 1.10.0

More info: https://github.com/google/googletest/releases/tag/release-1.10.0

* Adjust cmake variables to be used by Ninja of build byproducts

* Replace deprecated functions for googletest 1.10 release

* Add missing gtest_main target

* Fix locations for gtest and gtest_main targets

* Update CMake configuration for gtest tests

Tests were failing with next error: /usr/bin/ld: ../../../googletest-prefix/src/googletest-build/lib/libgtest_main.a(gtest_main.cc.o): in function testing::InitGoogleTest(int*, char**)
/usr/bin/ld: gtest_main.cc:(.text.startup+0x2f): undefined reference to testing::UnitTest::Run()

* Patch by Bertrand Bellenot: Update build configuration for Windows

* We build googletest in Release mode (for Debug names of libraries should be different and require special treatment)

* [formating] Add EOL in CMakeLists.txt

The weak vtable forces to compiler to duplicate the vtable for every
TU which includes the header and the deserializer to deserialize the
entries from PCH/PCM at startup.

One could Enable Console Virtual Terminal Sequences (ANSI escape code) that can control cursor movement, color/font mode, and other operations when written to the output stream, but then it breaks the wrap at eol output. This is a known issue, see https://github.com/microsoft/terminal/issues/349. With this feature, the escape sequences like `\033[39m` would work in the Windows 10 command prompt as well.

  - TMVA_Higgs_Classification.C

   a tutorial using the public Higgs UCI dataset for a classification problem using a Deep Neural network from TMVA, that is  made with fully connected layers

 - TMVA_CNN_Classification.C

 Tutorial showing the usage of Convolutional neural network in TMVA.
 The macro generates on the fly some toys images (size 16x16) of two different classes and then a
 convolutional neural network is used for their classification.
 This example builds and uses also a CNN built on the fly using Keras through the ROOT PyKeras package
 This example shows also how to use a batch normalization layer in TMVA

- TMVA_RNN_Classification.C

 Tutorial showing the usage of Recurrent neural network in TMVA.
 Toys time dependednt data of two different classes are generated on the fly and then
 a recurrent neural network is used for classification.
 Both TMVA and PyKeras networks are built and used.
 The network uses by default one LSTM layer, but optionally it can be built with
 a simple RNN or a GRU layer or also 3 different recurrent networks for each recurrent layer type
 can be made

Try to exclude as much as possible ${CMAKE_BUILD_DIR)/include from includes paths.
By this one much better control library dependencies - includes from other ROOT libraries not "visible" during the build.

Several generated files placed first to ${CMAKE_BUILD_DIR)/ginclude and then copied to include.

Dictionary generation still uses only ${CMAKE_BUILD_DIR)/include

Co-authored-by: Axel Naumann <Axel.Naumann@cern.ch>

 Do not build cudnn full test for simple RNN when cudnn is not available

 Use a single test program code ( RNN/testFullRNN.h) and remove teh ones in LSTM and GRU

- Fix a problem found in updating weights using gradient and learning rate in case of Cudnn

- remove an empty file (TestRecurrentForwardPassCuda.cxx)

 - implement the backpropagation for CPU for teh GRU in case the option reset gate after is on, this means that the
   candidate gate output is computed as   :        c(t) =  tanh ( W * x(t) + r(t) * ( U * h(t-1) ) + B )
  instead of the vanilla GRU implmentation :       c(t) =   tanh ( W * x(t) + U * ( r(t) * h(t-1) ) + B )

  The Cuddnn implementation of GRU is not the vanilla GRU as described for example in
   https://en.wikipedia.org/wiki/Gated_recurrent_unit#Fully_gated_unit

When computing the candidate gate the multiplication of the candidate weight state is done only with the previous state and the
multiplication with the reset state is done afterwards.
See  https://docs.nvidia.com/deeplearning/sdk/cudnn-api/index.html#cudnnRNNMode_t
We use in Cudnn this mode with a single input bias

- Ad test for comparing CPU and GPU implementations for all recurrent layer types

This commit fixes failures observed in incremental builds with debug mode

- Improve LSTM abd GRU Layer to add the option to return as output the full time sequence (statesize x time size) instead of
only the  last output (default case). This is needed for example when chaining RNN layers.

- CHave a single parsing function for Recurrent layer in the MethodDL to avoid code repetition

- Fix and improve all backpropagation tests for different configuration. Few bugs in backpropagation test programs for all 3 recurrent layers  have been fixed

Add the fast_tanh vdt implementation as a new activation function (FTANH) and do not use anymore as default implementation for tanh for the CPU architecture.