Modernize extension-cpp; refactor code

This PR:
- creates a single unified build for extension-cpp (instead of having
  separate cpu/cuda setup.pys).
- Updates the build system to use pyproject.toml (instead of only
  setup.py)
- Uses TORCH_LIBRARY to bind operators (instead of using PyBind)

There will be more future work to add improvements (e.g. torch.compile
support) and also fix up the corresponding C++ extensions tutorial.

Test Plan:
- Refactored all of the tests under test/

ghstack-source-id: a31d059ed3f83118513791e712a394713e07a96b
Pull Request resolved: #86

README.md

@@ -1,16 +1,24 @@
 # C++/CUDA Extensions in PyTorch
 
-An example of writing a C++ extension for PyTorch. See
+An example of writing a C++/CUDA extension for PyTorch. See
 [here](http://pytorch.org/tutorials/advanced/cpp_extension.html) for the accompanying tutorial.
+This repo demonstrates how to write an example `extension_cpp.ops.lltm`
+custom op that has both custom CPU and CUDA kernels.
 
-There are a few "sights" you can metaphorically visit in this repository:
+To build:
+```
+pip install .
+```
 
-- Inspect the C++ and CUDA extensions in the `cpp/` and `cuda/` folders,
-- Build C++ and/or CUDA extensions by going into the `cpp/` or `cuda/` folder and executing `python setup.py install`,
-- JIT-compile C++ and/or CUDA extensions by going into the `cpp/` or `cuda/` folder and calling `python jit.py`, which will JIT-compile the extension and load it,
-- Benchmark Python vs. C++ vs. CUDA by running `python benchmark.py {py, cpp, cuda} [--cuda]`,
-- Run gradient checks on the code by running `python grad_check.py {py, cpp, cuda} [--cuda]`.
-- Run output checks on the code by running `python check.py {forward, backward} [--cuda]`.
+To test:
+```
+python test/test_extension.py
+```
+
+To benchmark Python vs. C++ vs. CUDA:
+```
+python test/benchmark.py
+```
 
 ## Authors