[DeepSeek V4][NPU] Support CAMP2pAFDConnector

[jiangkuaixue123]

Parent: #169

Goal

Add DeepSeek V4 support on NPU through the synchronous CAMP2pAFDConnector path.

Priority

Follow-up — schedule after the first priority group (GPU PyNCCL and NPU async connector support).

Scope

• Integrate the DeepSeek V4 role-aware model adapter with CAMP2pAFDConnector.
• Validate hidden-state, residual, routing/gate, auxiliary payload, DP metadata, and TP/DP rank mapping contracts.
• Validate CAMP2P/HCCL shapes, dtypes, topology, handles, transfer state, supported eager/ACL graph/DBO-uBatch behavior, profiling, graph replay, failure cleanup, and weight-loading contracts.
• Add focused unit/contract tests and DeepSeek V4 NPU synchronous E2E coverage.
• Run the full affected NPU E2E regression suite.
• Publish a reproducible recipe under recipe/npu/CAMP2pAFDConnector/deepseek_v4/.

Dependencies

Implementation starts after #168 is complete and the NPU software stack is frozen as specified in #169.

Acceptance criteria

• DeepSeek V4 runs through CAMP2pAFDConnector on the frozen [RFC]: Upgrade the NPU backend to a vLLM-Ascend nightly compatible with vLLM 0.26.0 #168 stack.
• Correctness/accuracy is validated against native colocated vLLM-Ascend.
• Supported execution modes and TP/DP topologies have recorded evidence.
• Existing NPU E2E coverage remains green.
• Unsupported combinations fail early and are documented.
• The connector-specific recipe is complete and reproducible.