add warp transducer

[vincentqb]

This pull request introduces rnnt_loss and RNNTLoss in torchaudio.prototype.transducer using HawkAaron's warp-transducer.

• The python interface remains currently the same as the original.
• Update cmdclass to use custom clean here.
• The library is compiled statically, e.g. enable static build vincentqb/warp-transducer#2.
• Add a unit test for CPU.

Import the warp-transducer repository

Suggested options to load the repository with torchbind replacing pybind:

• (current, load as a submodule with bindings and CMakeList copied) Load original repository and bind outside of repository.
• Load original repository and patch.
• Load a fork like replace pybind by torchbind vincentqb/warp-transducer#1 as a submodule.
• git subtree as suggested in comment.
• Submit an upstream patch

CUDA

This pull request can optionally build with GPU support. Environment variables need to be set before installing, as mentioned here and colab notebook.

export CUDA_HOME=/usr/local/cuda
export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME
export LD_LIBRARY_PATH="$CUDA_HOME/extras/CUPTI/lib64:$LD_LIBRARY_PATH"
export LIBRARY_PATH=$CUDA_HOME/lib64:$LIBRARY_PATH
export LD_LIBRARY_PATH=$CUDA_HOME/lib64:$LD_LIBRARY_PATH
export CFLAGS="-I$CUDA_HOME/include $CFLAGS"
BUILD_CUDA_WT=ON python setup.py install

There is also a a unit test available for GPU.

cc @astaff, internal

[cpuhrsch]

Why is this needed?

[cpuhrsch]

I'd make this switch dependent on whether CUDA is available at all.