Merged commit includes the following changes:

228190834  by A. Unique TensorFlower<gardener@tensorflow.org>:

    Internal change

--
228187607  by A. Unique TensorFlower<gardener@tensorflow.org>:

    Split depthwise conv GPU code into multiple files

    This file was a bottleneck during compilation, often taking many minutes to compile. In local testing this change reduces the wall-clock build time for the depthwise conv GPU kernels from 163s to 105s.

--
228181904  by A. Unique TensorFlower<gardener@tensorflow.org>:

    Fix py2 and py3 of gen_git_version.py to fix 1.13 release.

    Byteify function to handle py2 and py3.

--
228161227  by A. Unique TensorFlower<gardener@tensorflow.org>:

    Create Tensorflow version of LLVM's opt.

    This is useful for iterating on individual passes in the OptimizationPassRegistry.
    This is run with a command of the form optimization_pass_runner --input_file_path=/tmp/input.pbtxt --output_file_path=/tmp/output.pbtxt --optimization_pass=NameOfGraphOptimizationPass

--

PiperOrigin-RevId: 228190834