Add MAGI-2 Preview support and MUSA optimizations

[yeahdongcn]

Summary

• Add MAGI-2 Preview diffusion pipeline support on top of upstream main.
• Add MUSA/S5000 execution path with MATE FlashAttention, SGL fused MoE, fused SwiGLU7, mHC fused path, owned W13/W2 layouts, and fast RoPE.
• Keep the implementation focused on code; benchmark details are tracked separately in the workspace ticket.

Validation

Workload: 272p/10s, 256x448, 125 frames, seed 42, 100 inference steps, EP4 + Ulysses8.

• MTT S5000 host 10.20.35.88: steady mean 781.162 ms/DiT step, p50 780.776 ms, p90 781.880 ms, max 815.733 ms.
• Generated 10-second MP4 was validated with ffprobe; checksum and full environment are tracked in workspace evidence.
• MAGI-2 Python modules compile successfully from the upstream-main worktree.

The historical 754.388 ms result is retained only as unverified evidence because its complete environment header was not captured; the number above is the conservative reproducible result.

Notes

This is a draft PR for review. Optimization commits were rebased onto the latest upstream main; follow-up splitting can be done after API review.

[hsliuustc0106]

This PR touches vllm_omni/diffusion/, tests/diffusion/, docs/design/, vllm_omni/model_extras/, tests/e2e/ (35 files). Based on CODEOWNERS coverage of the changed files, the most-related reviewers appear to be:

@wtomin @Bounty-hunter @fhfuih

Could one of you take a look when you get a chance? Thanks!