[PyTorch] Debug dtype casting in operation-based API

[timmoon10]

Description

Some users initialize model dtypes using module casting functions like torch.nn.Module.float. However, we experience weird errors when casting fusible operations, largely because the module casting functions don't properly handle Float8Tensor. This PR monkey-patches some module casting functions (float, half, bfloat16) so they properly handle Float8Tensor. This PR also adds logic in the linear operation to respect PyTorch autocasting.

Type of change

• [ ] Documentation change (change only to the documentation, either a fix or a new content)
• [x] Bug fix (non-breaking change which fixes an issue)
• [ ] 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

• Support setting Float8Tensor.data to arbitrary tensors
• Monkey-patch module dtype cast functions (e.g. torch.nn.Module.float) to handle Float8Tensor
• Respect PyTorch autocast dtype in linear 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

Note: L40 test failure disappears when job is rerun.

[timmoon10]

/te-ci pytorch

[timmoon10]

/te-ci pytorch

[timmoon10]

/te-ci pytorch