[PyTorch] Adjust checkpointing of FP8 metadata for attention

[cyanguwa]

Description

This PR relocates the FP8 metadata for attention from FusedAttention to DotProductAttention. It makes DotProductAttention a TransformerEngineBaseModule and FusedAttention a torch.nn.module. In the future, core_attention._extra_state will be the centralized place for FP8 metadata for any attention backend, instead of core_attention.fused_attention._extra_state which was just for FusedAttention (originated from #768 ).

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)

Changes

Please list the changes introduced in this PR:

• subclassed DotProductAttention to TransformerEngineBaseModule
• reverted FusedAttention to torch.nn.module
• passed fp8_meta from DotProductAttention to FusedAttention
• added miscellaneous warnings and tests

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
• [x ] 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

[cyanguwa]

/te-ci pytorch

[cyanguwa]

/te-ci pytorch

[timmoon10]

I think this approach is hacky. It adds an unnecessary fp8_meta to DotProductAttention and it intertwines the implementations of DotProductAttention and FusedAttention. I think the right solution is to refactor FusedAttention so it's not a subclass of TransformerEngineBaseModule. That way the fp8_meta lives in DotProductAttention, which passes it to FusedAttention.

I won't block since we want to get a bugfix before the release. It doesn't affect the checkpoint format, so we can fix the implementation in the future.

[cyanguwa]

/te-ci pytorch

[timmoon10]

LGTM