fix(hcu): avoid double router weighting in DeepEP GEMM

[alexanderbin123]

Summary

• align HCU DeepGEMM and DeepEP HT reduction with the official vLLM apply_router_weight_on_input contract
• apply router weights exactly once: at input or during unpermute/reduce
• share the reduction-weight helper across native DeepGEMM, Channel-FP8/INT8 DeepEP HT, and SlimQuant W4A8 HT
• keep DeepEP LL masked/N32, pure-TP AITER, and other quantization ownership boundaries unchanged

Reference

This follows the official vLLM modular MoE design:

• prepare/dispatch may apply router weights to the input
• expert execution produces unweighted expert results
• finalize/unpermute/reduce must use uniform weights when input weighting already occurred

Reference repository: https://github.com/vllm-project/vllm

Validation

Passed before opening this MR:

• tests/runtime_patch/test_deep_gemm_utils.py: 13 passed
• tests/runtime_patch/test_quant_gemm_aiter.py: 195 passed
• tests/runtime_patch/test_moe_deepep.py: 84 passed
• Python AST/bytecode compilation for modified modules
• git diff --check
• Qwen3.5-9B and Qwen3.5-35B-A3B integration test collection executed successfully

Additional real-model validation on 2026-09-03 using 8 gfx936 devices:

• Qwen3.5-35B-A3B-W8A8, TP8/EP8 smoke: passed with allgather_reducescatter
• Qwen3.5-35B-A3B-W8A8, DP8/EP8 with explicit DeepEP high-throughput: passed; selected DeepEPHTPrepareAndFinalize
• Qwen3.5-35B-A3B BF16, DP8/EP8 with explicit DeepEP high-throughput: passed; selected DeepEPHTPrepareAndFinalize
• Qwen3.5-35B-A3B BF16, DP8/EP8 with explicit DeepEP low-latency: passed; selected DeepEPLLPrepareAndFinalize

Environment-limited or unsupported configurations:

• Qwen3.5-9B smoke: skipped because the HCU model runtime resource was unavailable
• Qwen3.5-9B graph parity: skipped because the HCU model runtime resource was unavailable
• Qwen3.5-35B-A3B TP+EP matrix: 6 skipped because the test requires gfx938 and the available devices are gfx936
• Qwen3.5-35B-A3B-W8A8 with deepep_auto: rejected by the existing HCU guard because this Channel-INT8 auto path is validated only for DeepSeek-V4; explicit DeepEP HT is the supported Qwen3.5 route
• Qwen3.5-35B-A3B-W8A8 with explicit DeepEP LL: rejected because no Int8 MoE backend supports this deployment configuration
• Qwen3.5-35B-A3B-W8A8: no dedicated upstream local integration test was present in this v0.25.1 test tree

Known unrelated test prerequisite:

• the full test_moe_deepep.py module has one existing environment failure in test_moe_runner_and_shared_experts_cold_replacement_contract because /models/vllm_0251 is not available; the remaining 84 tests passed.