MacOS build issue with OpenMP

[mthrok]

Since around May 12th, the macOS build process stopped working on CI.

https://hud.pytorch.org/pytorch/audio/commit/f5036c7182f43e06ddf12f2b327cafaecf5763cf

Note that the couple of commits made before the issue started were Python changes, so it's unlikely that these commits broke the build process.

The cause is CMake not being able to find OpenMP. This has been working for quite a while.

In #2404, we install libomp from brew and this appeared to work, but the unittest seg-faulted on exact function that uses OpenMP.

https://app.circleci.com/pipelines/github/pytorch/audio/10846/workflows/3ae28b1a-e2a9-46c5-80d5-44aa95aad869/jobs/674356

This has been worked around by disabling OpenMP for macOS in #2431 .

The root cause of why it started is still unknown and not sure if this keeps us from re-enabling OpenMP for macOS.

The underlying issue is that there is no proper/clean way to detect/use the libomp that PyTorch was compiled with.
The following logic is used to enable/disable the OpenMP, but the success seems to depend on if the build environment has the correct/matching libomp that PyToroch used.

audio/tools/setup_helpers/extension.py

Line 43 in 6e56383

_USE_OPENMP = _get_build("USE_OPENMP", True) and "ATen parallel backend: OpenMP" in torch.__config__.parallel_info()

cc @atalman