perf(npu): optimize K3 dense/static verify path on A5

[McZyWu]

Summary

This is a new PR against a5-k3-0828 that carries the K3 KDA/DSpark
hot-path port from #8 and adds the dense/static follow-up optimizations.

• Keep the KDA raw/preactivated gate contract explicit, including K3
  lower_bound handling.
• Reuse fixed verify metadata and the existing fused KDA verify pieces.
• Add an opt-in fixed-width [B, T, C] causal-conv path for dense/static
  target verify (SGLANG_NPU_KDA_DENSE_CONV3D=1).
• Derive one graph-replay-safe int64 cache-index tensor from
  query_start_loc; zero-length graph-padding requests are mapped to -1
  and the same tensor is shared by causal-conv and KDA recurrent.
• Capture generic DSpark draft graphs at TP-local request batch sizes,
  including B1, while preserving TBO and context-parallel alignment.
• Keep dense output norm on the existing graph-proven fused path; the
  separate dense-only onorm experiment was not kept because it had no
  measurable graph gain.
• Add focused tests for DSpark graph-batch alignment and dense KDA padding
  metadata, plus a reproducible full-model TP32/DP1 launcher.

The formatting-only baseline fixes are isolated in commit 313acfa5f and
match the A5 branch's pre-commit run --all-files requirements.

Paired kernel PR

Requires zhaozx-cn/sgl-kernel-npu#3
from branch McZyWu:a5-k3-0828-dense-static-opt-20260903.

The framework passes -1 for graph-padding KDA state rows. The paired
kernel skips recurrence for that sentinel, zeroes the dense output row, and
does not touch the snapshot or persistent state.

Validation

• Framework unit tests in the CANN 9.0 container: 9 passed.
• Paired kernel target-verify and dense3d causal-conv tests: passed on NPU.
• Four-node kernel smoke tests on 209/212/216/217: 2 passed per node.
• Dense causal-conv microbenchmark (B=4,T=8,C=1152,W=4,state=10):
  20.387 us -> 18.446 us (1.105x) in the representative run; repeats
  were 1.078x-1.104x.
• Full 93-layer TP32/DP1 current-head run completed target graph capture on
  all 32 ranks (bs=[4,32]) and entered generic draft capture with
  bs=[1,4,32]. 29/32 ranks logged draft-capture completion before an
  external Docker SIGKILL stopped all four shared containers; the other
  three were interrupted in that same second. Dense3d was enabled and no
  KDA shape check, Triton BS conflict, traceback, or assertion occurred.
  End-to-end HTTP replay is pending an idle window.

The full run used SGLANG_RAGGED_VERIFY_MODE=static. Therefore the claimed
dense-path benefit comes from dense3d causal-conv, graph-safe padding, and
local draft graph buckets; compact-only ragged scatter/gather is not counted
as an active optimization in this result.

---

CI States

Latest PR Test (Base): ❌ Run #33762654951
Latest PR Test (Extra): ❌ Run #33762654713
Latest PR Test (AMD ROCm 7.2): ❌ Run #33762654653

[qybnb]

2ms