[Disagg] Support static DSpark PD for DeepSeek V4.1

[hnyls2002]

Summary

• Stack on dsv4.1: remaining model and runtime integration #38798 (dsv4.1) and enable DeepSeek V4.1 + DSpark + PD in static verify mode, including stochastic sampling.
• Require Mooncake, matching TP and DSpark block size, and PP/DP/CP size 1. Validate target/draft transfer layouts through Python and Rust bootstrap servers before caching peer mappings or transferring KV.
• Populate compression metadata on both workers and reuse the existing target/draft SWA and request-scoped pair-ring transfer.
• Route decode retraction through the existing prefill rebootstrap path, preserving the emitted boundary token and logprob instead of attempting unsupported CPU KV backup.
• Keep focused regression coverage for bootstrap compatibility and retraction boundary-token replay, plus compression metadata assertions in the existing wire tests.

Validation

• After pruning tests: 3 retained DSpark PD unit tests and 28 existing PD wire tests passed, along with pre-commit checks. The unchanged Rust implementation previously passed all 5 bootstrap tests, workspace clippy, and formatting on this upstream port.
• Before this port, on the official deepseek-ai/DeepSeek-V4.1-Flash@517ef625 checkpoint: full sgl-eval GSM8K 1,272/1,319 (96.4367%), zero errors/truncations/empty outputs. H200 TP4 prefill + TP4 decode, static block 5, non-thinking, temperature 0.7/top-p 0.9/seed 0, concurrency 8, max 2,048 output tokens. This used the support patch on the newer development base with official HF config support; no matched colocated GSM8K control was run.
• Before this port, on the earlier checkpoint: B300 AIME 2025 472/480 (98.3333%) PD versus 464/480 (96.6667%) colocated DSpark, 30 questions x 16 repeats, zero errors/truncations/empty answers. Matched thinking-mode sampling settings, TP4/EP4, static block 5; no forced retractions. H200 forced-retraction regression also passed with 32 x 512 sampled tokens and 133 retractions.
• The GPU results above remain historical validation; their diagnostic scripts and artifacts are retained outside this PR. GPU evaluations have not been rerun on this upstream commit. Cross-node RDMA and Rust frontend GPU E2E remain untested.

Known limitations

• Keep draft: the existing rebootstrap payload omits multimodal conditioning, and its internal HTTP request does not forward worker API-key authentication. The accuracy runs cover text-only requests without worker API-key authentication.
• Shape-dependent numerical differences also reproduce with PD and DSpark disabled. No production numerical fix is included; answer scores do not establish bitwise equivalence.
• No new steady-state decode kernel or synchronization; bootstrap validation is cached per peer. Retraction adds a device drain and prefix recomputation. No throughput A/B or zero-overhead claim.

---

CI States

Latest PR Test (Base): ❌ Run #34536354018
Latest PR Test (Extra): ❌ Run #34536353731
Latest PR Test (AMD ROCm 10): ❌ Run #34536354102