[PyTorch] Release GIL in PyTorch extensions

[timmoon10]

Description

This PR releases the GIL within PyTorch extensions to enable some multithreaded workflows (see https://github.com/NVIDIA/TransformerEngine/issues/868). As far as I'm aware of, the only time we call Python from within C++ is when initializing Userbuffers (i.e. when using NCCL to bootstrap UB comms).

Closes https://github.com/NVIDIA/TransformerEngine/issues/868.

Type of change

• [ ] Documentation change (change only to the documentation, either a fix or a new content)
• [ ] Bug fix (non-breaking change which fixes an issue)
• [x] New feature (non-breaking change which adds functionality)
• [ ] Breaking change (fix or feature that would cause existing functionality to not work as expected)
• [ ] Infra/Build change
• [ ] Code refractor

Changes

• Release GIL in PyTorch extensions

Checklist:

• [x] I have read and followed the contributing guidelines
• [x] The functionality is complete
• [x] I have commented my code, particularly in hard-to-understand areas
• [ ] I have made corresponding changes to the documentation
• [ ] My changes generate no new warnings
• [ ] I have added tests that prove my fix is effective or that my feature works
• [ ] New and existing unit tests pass locally with my changes

[timmoon10]

/te-ci pytorch

[denera]

Is there a way to test this?

We can manually validate that comm+GEMM overlap still works with this PR by running examples/pytorch/comm_gemm_overlap/ln_mlp_with_overlap.py. Everything else should be safe because they don't do any Python calls from C.