feat: extend DCP to GQA and optimize query communication

[CUHKSZzxy]

Motivation

Depends on #4952; merge that PR first. This extends its MLA DCP support to GQA models with replicated KV heads and adds an optional optimized query gather. The diff against main includes #4952 until it is merged.

Modification

• Add GQA DCP prefill/decode with bounded cached-prefix gathering and shared DCP components.
• Support unquantized and per-tensor FP8 KV cache, including sharded FP8 cache fill. Propagate the allocated draft cache's quantization policy into MTP attention to fix FP8 warmup failure.
• Add communicator-owned query gathering with opt-in symmetric memory (LMDEPLOY_ENABLE_SYMM_MEM_DCP=1) and a process-group fallback.
• Share KV workspace sizing and document supported layouts and restrictions.

Accuracy

Full GSM8K test split, zero-shot, thinking disabled, temperature 0, seed 20260921, output limit 2,048 tokens, concurrency 16. Both runs used FP8 KV and prefix caching on H200 GPUs.

Model	Configuration	Correct / total	Accuracy	Request / parser errors	Truncated
Qwen3-235B-A22B	TP8+DCP2	1253/1319	95.00%	0 / 0	1
Qwen3.5-397B-A17B-FP8	TP4+DCP2, MTP2, batch limit 128	1276/1319	96.74%	0 / 0	10

Both mini checks passed 8/8. Qwen3.5 MTP acceptance was 92.71%. Unspecified engine settings used CLI defaults. Qwen3 was tested before the final draft-policy fix; Qwen3.5 includes it. These are single runs without a TP-only baseline; prefix caching was enabled but no cache hits were observed. The optional symmetric-memory gather was not enabled in these accuracy runs.

Checklist

• Pre-commit checks passed for the final fix.
• Focused speculative-decoding tests: 24 passed, one CUDA-only test skipped.
• Context-parallel documentation updated.
• Full kernel/backend and dependency-version matrix remains unvalidated.

Assistance

Assisted with Codex + GPT-6-Astra High, reviewed manually