feat(npu): add DeepSeek V4 async CAM W4A8 layered FFN path

[jiangkuaixue123]

Purpose

Add an opt-in DeepSeek V4 W4A8 path for Async CAM FFN that keeps received layer IDs and routed token counts on device and calls layered W13/W2 GMM. The existing path remains the default.

Scope and implementation

• Ascend 910C, DeepSeek V4 W4A8, Attention-side gate, dynamicQuant=1, eager ModelRunnerV1, static routed experts, and per-channel or per-group loaded weights.
• AFD_ASYNC_CAM_LAYERED_GMM=1 enables the FFN path. Other roles and unsupported configurations fail early.
• Build per-layer weight and scale TensorLists after model load; map global MoE layer IDs to compact device indices. Materialize a one-element layer-index slice so CANN sees standalone storage.
• Run layered W13 and W2 on capacity-sized activations, then send compact CAM metadata to combine. No layer ID or token-count host read is introduced in the execution path.
• Adapt the vendored GMM op API to the checkpoint loader's logical [E,K,N] INT32 view over five-dimensional NZ INT8-packed storage. The change is limited to shape validation and packed storage interpretation; the compute kernel is unchanged.
• The model specifies swiglu_limit=10.0, but the current layered fused GMM does not implement that clamp. The opt-in path temporarily ignores this limit, logs that decision, and is not numerically equivalent by contract. Full-model accuracy was checked below; implementing the clamp remains follow-up work.
• Builds on the device group_list interface from fix(npu): accept device group lists in layered GMM #384.

Validation

• Pre-commit and DCO checks pass; targeted CPU/unit tests pass.
• CANN extension rebuilt on Ascend 910C. Synthetic layered W4A8 tests covered contiguous and noncontiguous layer IDs, NZ-packed weights, per-channel/per-group scales, and capacity tails. One full-capacity comparison fluctuated on its first run and passed on targeted rerun; this is retained as a numerical validation caveat.
• Prefill-only Attention DP2/TP4 + FFN EP8 with the layered path returned a correct chat response. GSM8K 5-shot, 64 concurrent, 300 samples: strict/flexible exact match 97.00% (291/300).
• PD deployment: Prefill Attention DP2/TP2 + FFN EP4 with layered W4A8; ordinary Decode DP8/TP1/EP8. End-to-end chat via the PD proxy returned HTTP 200 and the correct answer.
• PD GSM8K 5-shot, 64 concurrent, all 1319 test samples: strict/flexible exact match 95.15% (1255/1319); no request errors. The earlier PD configuration scored 95.22% strict and 95.30% flexible on the same 1319 samples. This is a model-level accuracy check, not proof of exact SwiGLU-limit equivalence.
• Decode must use the system CAM operator directory. Inheriting this PR's custom CAM directory caused the native Decode W4A8 fused GMM to fail shape validation during startup; restoring the system directory resolved it. This is a deployment environment setting, not a code change in this PR.
• The Async CAM FFN receive kernel can time out after roughly ten minutes without requests. The validation deployment uses a four-minute keepalive; idle-time handling remains open.
• No performance improvement is claimed yet.

Docs

Updated docs/npu/CAM_ASYNC_CONNECTOR_USER_GUIDE.md and docs/npu/TESTING.md for the opt-in path, scope, and validation.

---

Essential PR Checklist

• Purpose and supported scope are clear.
• vLLM v0.26.0 compatibility is considered.
• Imports remain CPU-safe; NPU operator loading is delayed until a qualified FFN path starts.
• CPU, NPU, deployment, and accuracy evidence is included.
• Numerical and idle-time limitations are documented.
• Documentation impact is stated.