feat(fastwam): add structured FlashAttention backend

[XueSongTap]

Summary

• Add structured attention execution for FastWAM MoT masks.
• Support FA4, FA3, FA2, SDPA, and automatic backend fallback.
• Integrate the backend across base MoT, Joint, and IDM paths.
• Preserve dense-mask semantics while avoiding dense [B, H, Q, K] masks.

Validation

• Added forward/backward equivalence tests for structured vs dense attention.
• Added mask-semantic tests for Base, Joint, and IDM variants.
• compileall and git diff --check pass.
• BF16 CUDA equivalence probe: forward max error 0.0, gradient max error 0.015625.

Performance

Microbenchmark (video tokens=588, action tokens=32, heads=24, head_dim=128, mbs16):

• Forward: SDPA 0.991 ms -> FA2 0.577 ms (1.72x).
• Forward + backward: SDPA 6.027 ms -> FA2 3.993 ms (1.51x).

End-to-end 8-GPU training loss remains aligned with the original FastWAM baseline; the current workload is communication-bound, so overall throughput is approximately unchanged.