[1/2] Wean off of PYBIND in favor of torch.ops.load_library

[janeyx99]

The no ghstack version of #1275 to see if the mac tests pass (they do).

Pave the path for python agnostic ao by removing depending on PYBIND (which is not python agnostic).

Concretely, what happened this PR?

• no more PYBIND, so no more init.cpp
• for all non-CUDA platforms, ao no longer has custom cpp extensions, and thus ao is a pure python lib
• so we skip auditwheel (which is used for when the platform is linux) for all non-cuda wheel builds. (This does also mean we now needlessly build the same python wheel for every other platform rocm, cpu, xpu...)
• no more PYBIND also means no more torchao._C, which we're replacing with a load_library call

This PR should have no failures. The next PR will be targeting the wheel process to only output one wheel for every python version. If you're curious, the next PR looks like #1277

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/1276

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit acd19e0 with merge base f96e5ec ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[msaroufim]

is this also true for rocm?

[janeyx99]

No, but it looks like the current .cu files are not built for rocm based on the use_cuda here https://github.com/pytorch/ao/blob/main/setup.py#L64

[msaroufim]

Ok! heads up to @petrex who has been working on adding custom hip kernel support

[petrex]

@janeyx99 This #1201 would build hip kernels.