[Spec][DSpark] Remove overlap scheduler host syncs for pure TP

[tanth47]

Summary

• Keep speculative sequence lengths and DSpark verify planning device-resident in the overlap path.
• Preserve the two-publication-back confidence relay without blocking the scheduler on the current forward.
• Use deterministic pure-TP graph-tier coordination instead of a per-step CPU/Gloo broadcast.
• Reset relay/planner generations on cache flush and keep source/non-source TP readiness safe.
• Keep cap-accept and eager compact verify lengths on device, and avoid D2H materialization in the DSV4 DSpark draft metadata path.

This follows the spec-v2 run-ahead design from #31468. It also removes the pure-TP per-step CPU tier synchronization used by #31195; DP-attention still keeps conservative CPU coordination and is explicitly out of scope.

Scope

• Pure TP / non-DP only.
• No SPS table tuning, target-accept policy, or dry-run policy changes.
• DP-attention host coordination remains follow-up work.

Validation

• Clean MI350 GPU suite before rebase: 55 passed, 18 subtests passed.
• Cumulative S1b/S2 MI350 suite with the same production delta: 143 passed, 35 subtests passed.
• Rebase range-diff is unchanged: ea3866819 = 5fb4eea25, 28ed43334 = 35789fec6.
• git diff --check and focused compile checks pass.

Before Ready

• Rerun the focused MI350 suite on exact rebased head 35789fec6.
• Add warmed fixed-seed UltraChat ~2K-ISL, exact-512-OSL c1 overlap ON/OFF evidence.
• Add c16 throughput, AR/AL, target-verify graph, folded accept/commit, and one-forward guard evidence.

Roadmap: #30734.

---

CI States

Latest PR Test (Base): ❌ Run #30141995501
Latest PR Test (Extra): ❌ Run #30141995451

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.