perf: optimize GLM-5.2 serving

[CUHKSZzxy]

Summary

Optimize GLM-5.2 FP8 serving across attention, MoE, projections, weight loading, and the output head:

• Use DeepGEMM for DSA index scoring.
• Extend compact blocked-FP8 MoE to low route density and fuse noaux tensor-core routing.
• Fuse Q/KV-A projections and shard LM-head vocabulary across TP ranks.
• Refresh dense FlashMLA metadata correctly during CUDA-graph replay.
• Load merged blocked-FP8 scales into TP-local destination sections.

Dependency

Depends on #4805; its five commits currently form the branch prefix. GLM-5.2 support PR #4737 has merged.

Performance

GLM-5.2-FP8 TP8 on 8 H200 GPUs, FP8 KV cache, MTP5, concurrency 16, and five-turn SWE-Smith:

• Two fresh-server runs: 803.60 and 808.51 tok/s; mean 806.06 tok/s.
• +12.86% versus the 714.22 tok/s DeepGEMM-index baseline.
• Both runs completed 80/80 requests successfully.

These are cumulative branch measurements and predate the TP-sharded LM head.

Validation

• Pre-LM-head full GSM8K: 1,267/1,319 (96.06%), exactly matching the reference, with no request or extraction failures.
• Targeted FlashMLA, cache, DeepGEMM, FP8 MoE, routing, projection, and NSA tests passed after rebase.
• Full TP8 model loading, CUDA-graph warmup, and repeated SWE runs completed successfully.

Assistance

Assisted with Codex + GPT-5.6-Sol xHigh, reviewed manually