[Bench][SM70] Add DFlash2 selector alignment diagnostics

[yangzhuxinyzx]

Purpose

Add explicit, default-off DFlash2 selector alignment capture and an offline analyzer for exact proposal-distribution research. The tooling records the fixed top16 IDs, unary logits, the 7x16x16 selector lattice, realized q rows, target top20 rows, and strict rejection outcomes only when VLLM_SPEC_DUMP_ALIGNMENT is enabled.

This PR does not change the production proposal policy and makes no acceptance or performance claim. It provides the evidence path needed to develop a later default-on optimization.

Safety

• Disabled diagnostics allocate no shadow lattice and perform no tensor copy; the dump function is gated at the call site.
• Capture remains rank-0, B1 and limit/step filtered under the existing sparse-rejection contract.
• Private artifact paths are not committed.
• Runtime admission depends on SM70, sampler contract and explicit diagnostics, not model/checkpoint identity.

Test Result

• Merged onto main@9368f49c4679ee8901288f56a329ccb639b2c657.
• CPU DFlash2/analyzer suite: 78 passed, 12 expected CUDA skips.
• All changed-file pre-commit hooks pass, including Ruff, mypy and markdown.
• Real GPU trace, held-out survivor and performance/acceptance claims remain future implementation work.

[yangzhuxinyzx]

审计结论：当前不进入 main，保留研究分支后关闭。

这是诊断/R&D 工具集，不是已经闭环的优化：PR 明确写明 GPU trace、端到端 acceptance、性能与 +0.5 tokens/round 结果全部 pending。它新增约 1,396 行，包括运行时 shadow lattice、同步 GPU->CPU dump、两套 benchmark/analyzer 与固定质量 harness，但没有真实 dump 上的 held-out 结果，更没有候选配置进入同条件 A/B。主线也已经有通用 VLLM_SPEC_DUMP_ALIGNMENT 基础设施，因此仅把 DFlash2 专用采集器并入生产路径，会先增加维护面而没有推进实现决策。

可拆分审计未发现应单独合并的 bugfix：所有 runtime 变更都只服务于该诊断，默认关闭本身不是性能/质量证据。最优后续是从该分支完成真实 trace -> held-out survivor -> exact-q 实现 -> V100 延迟/接受率/质量门禁，再以小型实现 PR 提交；离线分析脚本可留作该实验分支工具。当前关闭不会删除提交或证据。

[yangzhuxinyzx]

源码审计完成：该变更仅增加显式启用、默认关闭的 DFlash2 selector alignment 诊断与分析工具，不改变默认推理热路径。专项测试 78 passed / 12 skipped；PR 改动文件的 pre-commit 全部通过。GitHub 全仓 pre-commit 失败来自 227 个历史基线文件被格式化，和本 PR 的 8 个文件重叠为 0，因此判定为非阻断基线噪声。