[PyTorch] Remove fast param getter from modules

[timmoon10]

Description

#1191 added a fast accessor for params to work around CPU overheads from how PyTorch modules store params. This saves ~1 us per forward pass (out of a total ~100 us) in small benchmarks with the linear layer. However, we have experienced some problems with FSDP since it internally replaces the params with a flattened param and turns the original params into a plain tensor. The fast accessor was very hacky, and this edge case makes me think it's not worth the performance benefit.

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)
• 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

• Remove fast param accessor in TE modules

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• 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

[timmoon10]

/te-ci pytorch

[ksivaman]

LGTM