integrate new AITER mha v4 fmha entrypoint to xDiT

[jcaraban]

Motivation

Expose the low-precision MHA v4 kernels added by ROCm/aiter#4627 through xDiT.

Technical Details

• Add I8FP8, MXFP4, MXFP6, F4F4, and F6F4 attention backends.
• Delegate preprocessing and kernel dispatch to AITER's public mha_v4 API.
• Support unequal query and KV sequence lengths.
• Validate availability through the public AITER API.
• Reject unsupported causal attention explicitly.
• Document the new gfx950, head-dimension-128 backends.

Testing

• Added eager correctness checks against SDPA.
• Added torch.compile(fullgraph=True) coverage.
• Covered unequal sequence lengths, cross-attention, partial tiles, and causal rejection.
• All 20 focused tests pass on gfx950.

[avjves]

Why are these separate from the above check?

[jcaraban]

because mha_v4_mxfp8 will fail to import until ROCm/aiter#4967 merges, so if it was with the above block it would make the other MHA v4 backends fail, while those import fine in AITER main