[C/PyTorch] Add support for multi-latent attention (MLA)

[cyanguwa]

Description

This PR adds support for multi-latent attention (MLA) where head_dim_qk != head_dim_v.

This feature is only supported on the DotProductAttention level and by both UnfusedDotProductAttention backend and FusedAttention backend.

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

Please list the changes introduced in this PR:

• Add support for MLA

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

[cyanguwa]

/te-ci

[cyanguwa]

/te-ci

[cyanguwa]

/te-ci

[cyanguwa]

Some CI issue. Local tests for L0_jax_unittest using the CI container pass with no problem on A100/V100.

[zlsh80826]

LGTM!

[JiwenJ]

Hello, does this support qk != v head dim