Fix 4P/4D GLM NIAH: Wei combine() topk_ids and pair warmup with score

[basemam]

Summary

• Runtime Wei Sun combine() patch (e25bf182): MoRI finalize() uses original topk_ids, not dispatched ids. Required at EP32 (4P/4D). Runs even with GLM_SKIP_PATCHERS=1. Does not change P/D vllm serve argv (FULL graphs, decode mnbt=2048, kv-config, Rust vllm-router).
• NIAH client only: BENCHMARK_SCRIPT=niah, still POST /v1/chat/completions. Pairing is warmup N then score N (not warm-all-then-score-all).

Result — job 217082 (4P/4D, GLM-5.1-FP8)

Image rocmshared/pytorch-private:glm5.1-pr176-verify, GLM_SKIP_PATCHERS=1, Wei combine applied at serve, Rust router, NIAH_SEEDS=0, NIAH_HALT_ON_FAIL=0.

=== NIAH summary (mean/min/max across 1 seed(s)) ===
  words=  2000  mean=9.0/10  min=9  max=9  (n=1)
  words=  8000  mean=10.0/10  min=10  max=10  (n=1)
  words= 16000  mean=8.0/10  min=8  max=8  (n=1)
  words= 20000  mean=9.0/10  min=9  max=9  (n=1)
  words= 28000  mean=10.0/10  min=10  max=10  (n=1)
  words= 35000  mean=10.0/10  min=10  max=10  (n=1)

Log: /shared_inference/bbarakat/model_blog_logs/217082/niah_217082_20260818_195513_xP4_yD4_GLM-5.1-FP8.log