[PyTorch] Fused dbias-cast-transpose in bias operation

[timmoon10]

Description

This PR modifies the bias operation (see the operation-based API in https://github.com/NVIDIA/TransformerEngine/pull/707) so that it can launch a fused kernel for computing bgrad, FP8 dgrad, and FP8 dgrad transpose.

Once https://github.com/NVIDIA/TransformerEngine/pull/1164 is merged, this could be extended to call a fused kernel for dgelu, bgrad, FP8 dgrad, and FP8 dgrad transpose. However, it may be better to defer that to a followup PR since it will require some infrastructure changes (e.g. we want to apply different fusions in the FP8 and non-FP8 cases).

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

• Fused dbias-cast-transpose kernel in bias operation

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
• [x] I have made corresponding changes to the documentation
• [ ] My changes generate no new warnings
• [x] I have added tests that prove my fix is effective or that my feature works
• [x] New and existing unit tests pass locally with my changes

[timmoon10]

/te-ci pytorch

[timmoon10]

/te-ci pytorch

[timmoon10]

/te-ci pytorch

[timmoon10]

/te-ci pytorch