[PyTorch] Add max_score support for MuonClip

[cyanguwa]

Description

This PR adds max_score support in FusedAttention and UnfusedDotProductAttention backends in TE-PyTorch. The max_score can be used by a MuonClip optimizer. This PR only support FP16/BF16 and non-CP 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)
• 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 refactoring

Changes

Please list the changes introduced in this PR:

• Add max_score support in FusedAttention and UnfusedDotProductAttention

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

[skyw]

This changes returned values when max score is not requested. Should we keep backward compatibility if new feature is not used?