perf(gms/sglang): amortize safe TP lease coordination

[hutm]

Summary

Batch SGLang TP ownership coordination while preserving rollback and pressure safety.

Why this is needed

Per-page cross-rank synchronization is correct but can materially regress steady-state latency and throughput. Batching removes that hot-path cost without weakening ownership.

Validation

• The cumulative publication tip is byte-identical to the locally validated implementation.
• Cumulative focused validation: 994 passed, 1 skipped; the later vLLM-focused rerun passed 244 tests.
• Local TP2 crash tests passed exact-output recovery for vLLM and SGLang; final performance qualification remains tracked in DEP: Production hardening for persistent GMS KV failover #15035.
• This draft checkpoint still requires normal CI and reviewer validation before merge.

Stack

• Part 6 of 15 in the production-hardening train.
• Draft DEP and complete review order: DEP: Production hardening for persistent GMS KV failover #15035.
• Base: hutm/gms-hardening-05-acceptance-harness.

This PR is intentionally draft and stacked on its immediate predecessor.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.