[Cherry-pick to release/v0.5.15] Stabilize GLM-5.2 MTP IndexShare across PD and CUDA graph replay (#30839)

[kpham-sgl]

Backport of #30839 to release/v0.5.15.

Structured as two commits to stay conflict-free against the release branch:

1. Reapply [Spec] Anchor GLM-5.2 MTP IndexShare topk on the draft-extend step #29787 — reverts the release-branch revert Revert "[Spec] Anchor GLM-5.2 MTP IndexShare topk on the draft-extend… #30842, reinstating the GLM-5.2 MTP IndexShare draft-extend anchor that this fix builds on.
2. Squashed backport of [bug-fix] Stabilize GLM-5.2 MTP IndexShare across PD and CUDA graph replay #30839 — the stabilization work itself.

Motivation

This PR makes GLM-5.2 MTP IndexShare reliable across PD disaggregation, context parallelism, overlap scheduling, batch splitting, and CUDA graph replay. It preserves the correct draft-extend DSA seed across draft steps, avoiding redundant indexer computation without reusing stale or misaligned indices. This improves proposal consistency, acceptance length stability, and inference performance.

This is a follow-up to #29787. The optimization remains gated to EAGLE with speculative_eagle_topk == 1, index_share_for_mtp_iteration=true, and a valid index_topk.

Modifications

• Transfer per-request DSA top-k seeds from prefill nodes to decode nodes in PD disaggregation, with explicit invalid sentinels and retry-state cleanup.
• Unify eager and CUDA-graph seed capture through LogitsProcessorOutput, including correct last-token extraction after CP gather/reordering.
• Preserve seed alignment across context parallelism and Two-Batch Overlap splitting.
• Keep reusable seed state across copied forward batches and graph capture/replay.
• Track seed validity in FutureMap to prevent stale overlap/PD state, and fall back to eager recomputation when no valid seed is available.
• Refresh paged-MQA schedules, top-k-v2 plans, and per-step metadata during multi-step DSA graph replay.
• Resolve DSA context-parallel aliases from the effective attention backend.

🤖 Generated with Claude Code

---

CI States

Latest PR Test (Base): ❌ Run #29295651002
Latest PR Test (Extra): ❌ Run #29295650791

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[kpham-sgl]

Internal nightly run #0e8ae04b has passed GSM8K. Accept length of PD Disagg also looks healthy > 5

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[Jiminator]

Validated with GLM-5.2-NVFP4 (agentic multi-turn workload, ISL ≈ 80k, EAGLE 5/1/6) on 4×GB300 and 8×B300: no perf regression (within run-to-run noise vs release without this change), acceptance length unchanged, and no degenerate output across all generated turns.

4×GB300

Concurrency	TP=4 tok/s/gpu	TP=4 tok/s/user	TEP=4 tok/s/gpu	TEP=4 tok/s/user
1	15,452 *	455	17,728	439
2	25,062	265	25,359	258
4	32,074	156	32,239	146
8	37,886	78	38,006	80
16	40,921	39	42,089	41

8×B300

Concurrency	TP=8 tok/s/gpu	TP=8 tok/s/user	TEP=8 tok/s/gpu	TEP=8 tok/s/user
1	10,393 *	524	11,263	505
2	16,108	364	16,252	341
4	21,173	239	20,822	195
8	24,957	108	25,771	110
16	27,309	59	27,319	60

* depressed by one-time JIT compile on the cold first step; interactivity unaffected.