[PD] Fix TP-skew decode polling and Mooncake staging deadlocks

[YAMY1234]

Motivation / 动机

A low-concurrency disaggregated configuration (prefill TP1/DEP4, decode TP8, attention TP gather disabled) showed strong native batch-size-1 CUDA graph performance, but longer workloads could permanently stop making request progress.

低并发 PD 分离配置（prefill TP1/DEP4、decode TP8、关闭 attention TP gather）可以获得明显的原生 batch-size-1 CUDA graph 性能收益，但长时间 workload 会永久停止请求推进。

The hang came from three synchronization/state-machine bugs:

该 hang 由三个同步与状态机问题共同造成：

• Decode KV polling and request broadcast could enter different collective epochs across TP ranks.
• A deferred heterogeneous-TP Mooncake fanout could replay destinations that had already completed.
• An empty staging ring could retain a stale head, while a new watermark subscriber could miss the allocator's current state.

对应地：decode KV polling 与请求广播可能在不同 TP rank 上进入不同 collective epoch；异构 TP 的 Mooncake fanout 在 defer 后可能重放已完成 destination；空 staging ring 可能保留旧 head，而新 watermark subscriber 又可能错过 allocator 当前状态。

Changes / 修改

• Serialize decode polling and request broadcast through one ordered coordinator, using fixed-shape collective payloads and task-bound results.

• Track completed fanout destinations across requeue so staging transfers are idempotent.

• Reset an empty staging ring to a fresh round and bootstrap new subscribers with the current watermark.

• Add focused regression tests for collective ordering, fixed-shape polling, fanout retry, ring wrap, and subscriber bootstrap.

• 通过单一有序 coordinator 串行化 decode polling 与请求广播，并使用固定形状 collective payload 和任务绑定结果。

• 在 requeue 期间记录已完成的 fanout destination，保证 staging transfer 幂等。

• 空 ring 进入新 round 时重置状态，并用当前 watermark 初始化新 subscriber。

• 为 collective ordering、固定形状 polling、fanout retry、ring wrap 和 subscriber bootstrap 增加定向回归测试。

Validation / 验证

• Linux targeted tests: 18/18 passed.

• Exact long-duration workload: 807/807 requests, 0 errors, 0 cancellations, 3630.00 s.

• TTFT and inter-token-latency coverage: 100% / 100%.

• P50/P90 output-token throughput per user: 467.70/447.07 → 640.33/685.20 tok/s/user (+36.9% / +53.3%).

• No model forward or sampling math is changed.

• Linux 定向测试：18/18 通过。

• 精确长时间 workload：807/807 requests、0 errors、0 cancellations、3630.00 秒。

• TTFT 与 inter-token-latency coverage：100% / 100%。

• P50/P90 每用户输出 token 吞吐：467.70/447.07 → 640.33/685.20 tok/s/user（+36.9% / +53.3%）。

• 未修改模型 forward 或 sampling 数学逻辑。

---

CI States

Latest PR Test (Base): ⏳ Run #33196082438
Latest PR Test (Extra): ❌ Run #33196082037
Latest PR Test (AMD ROCm 7.2): ⏳ Run #33196083112

[YAMY1234]

/tag-and-rerun-ci