[Bugfix][Model][SM70] Repair Qwen3.8 Flash Next correctness

[yangzhuxinyzx]

Purpose

Repair the Qwen3.8-Flash-Next correctness gaps found while auditing current
1Cat public main against the live vLLM support PRs, without weakening the
existing V100/SM70 routes.

Base SHA: 62ad1e02693f4c857f3b7547cef1860ee54e8053.

This PR:

• adapts vLLM #53896's request-layout-safe PLE n-gram ID custom op while
  preserving 1Cat's pinned-host CPU PLE offload;
• adapts vLLM #53802/#54076 so hybrid Mamba/GDN prefill uses the real Mamba
  state grid and stops at every state boundary, including the (n - 1) tail;
• adapts vLLM #48375 so EAGLE/MTP drops the final Mamba cache-hit state;
• rejects the unvalidated legacy Qwen4Exp runner and keeps QSA from claiming
  batch-invariant split-K reductions;
• pins the packed GDN FP32-beta behavior with an exact V100 regression; and
• changes online QPN8 from default-on to explicit experimental opt-in. QPN8
  requantizes dense checkpoint weights at load time; the exact checkpoint
  NVFP4 QPN-M1 expert route remains default-on and unchanged.

Output-quality decision

The historical online-QPN8 A/B kept 96/96 GSM8K answers correct but only 8/96
token sequences matched. One answer expanded from 1,653 to 3,626 tokens and
its repeated-4-gram ratio rose from 0.199 to 0.507. This does not justify a
precision-preserving default, so VLLM_SM70_QWEN4_EXP_ONLINE_QPN8 now defaults
to 0 and requires explicit =1 opt-in with a quality warning.

The previous ~82.3 tokens/s result used online QPN8 and is not an accepted
no-loss baseline. The precision-preserving no-MTP control is about 67.6
tokens/s; reaching 80 tokens/s without QPN8 remains a separate optimization
target.

Test Plan

• Run focused Qwen4Exp configuration, PLE, QSA, hybrid-cache, prefix-cache,
  and QPN policy tests.
• Run the native QPN operator test against the existing source-built sidecar.
• Run packed GDN reference and exact FP32-beta tests on V100.
• Run all applicable pre-commit hooks on every changed file.
• When a clean TP4 window is available, run one combined V2/no-MTP/CPU-PLE
  model gate with QPN8 unset, rather than restarting once per scenario.

Test Result

• 124 passed, 1 deselected, 18 warnings in 52.17 seconds for the focused
  Qwen4Exp/hybrid/prefix/QPN suite.
• Native QPN sidecar operator: 1 passed with explicit online-QPN8 opt-in.
• Packed GDN: 7 passed, including the exact FP16-input/FP32-beta state test.
• Changed-file pre-commit: all applicable Ruff, format, typos, markdownlint,
  mypy, SPDX, forbidden-import, configuration, and API checks passed.
• git diff --check passed.
• The downloaded RadixArk index was checked against all 296,475 tensors in 206
  safetensors files: zero extra and zero missing keys, so vLLM #54230 does not
  affect this checkpoint.

No branch E2E result is claimed yet. GPUs 0-3 host the existing public MTP4
service, and GPUs 4-7 were held by a separate Nsight task at validation time;
this PR deliberately did not preempt either workload or repeatedly restart the
model.

Remaining gates and dependencies

• 1Cat [Bugfix] Avoid KVBlockZeroer H2D staging races #406 owns the independent KVBlockZeroer asynchronous H2D staging
  lifetime fix and should land before high-concurrency prefix-cache acceptance.
• vLLM issue #54199 remains open for overlapping donor/request lifetimes in
  Mamba prefix-cache precopy. A bounds-only workaround is not included because
  it can replace a crash with silent recurrent-state corruption.
• A final combined TP4/V2/no-MTP/CPU-PLE gate is required before merge. It must
  cover deterministic generation, tool calls, PLE request-layout changes, and
  prefix lengths around the real state boundaries.

Duplicate-work audit

• vLLM #51599's async accepted-count race is already covered by the local MRV2
  runner-owned snapshots and request-ID remapping.
• vLLM #53520 fixes legacy-runner prompt-logprob ordering; MRV2 computes prompt
  logprobs before its drafter, and this PR rejects Qwen4Exp legacy-runner use.
• vLLM #53122 targets standalone DFlash draft quantization; native Qwen4Exp MTP
  already calls configure_quant_config(Qwen4ExpMTP).
• vLLM issue #53982's circular block-table OOB path is already avoided in both
  local runners by disabling generic slot mapping for QSA circular groups.
• Disk/mmap PLE PRs are intentionally not imported because the target contract
  keeps PLE in host RAM.

AI assistance was used for upstream comparison, implementation, tests, and the
audit write-up. Every changed line and reported result was reviewed locally.