fix(vLLM): harden native IPC synchronization

[CjhHa1]

Summary

Hardens the native vLLM packed-IPC path added in #480 against rank-local failures and memory spikes.

• records the canonical outputs produced by each FSDP materialization and reaches rank consensus before the next materialization collective;
• also gates IPC-buffer allocation and packed-chunk handle gathering, so one rank cannot leave peers blocked in another process group;
• drops full-tensor references after their packed copy instead of retaining the previous tensor while materializing the next one;
• requires the Gloo control timeout to exceed the vLLM weight-update timeout, and fixes the TP4 reference recipe accordingly.

Related Issue

N/A

Test Plan

• python3 -m pytest -q tests/test_vllm_native_engine.py tests/test_ipc_weight_sync.py (8 passed)
• SKIP=no-commit-to-branch pre-commit run --all-files --show-diff-on-failure
• push-time pre-commit suite
• GPU smoke test not run; this environment has no allocated GPU. The focused tests cover source/chunk consensus gates, failure routing to the next safe gate, buffer-allocation consensus, boundary validation, and timeout ordering.

Compatibility / Risk

• No checkpoint or wire-format changes.
• Native vLLM IPC now performs a Gloo consensus at each raw state-dict materialization boundary; this adds control-plane synchronization to prevent cross-process-group deadlocks.
• Existing valid timeout configurations remain compatible. Native IPC configurations where control_timeout_s <= weight_update_timeout_s now fail during construction instead of timing out asymmetrically during publication.

Reviewer Notes

Follow-up to #480. Please focus on VLLMNativeIPCTrainerEngine._send_planned_packed: source boundaries guard the next FSDP collective, while chunk boundaries guard the default-group IPC-handle gather. The two gates are intentionally distinct.

Checklist

• I reviewed the changed code and removed unrelated/generated artifacts.
• I updated tests, docs, and configs where needed, or explained why not.