Feat: add weigh transfer fix

[S1ro1]

Note

Medium Risk
Patches vLLM DP engine core request handling and wave coordination logic, which is concurrency-sensitive and could impact distributed inference behavior if assumptions differ across vLLM versions.

Overview
Adds a new vLLM monkeypatch to prevent a pause/resume deadlock in DP mode by changing EngineCore.add_request behavior: it now only emits start_wave notifications when the scheduler is unpaused and forces engines_running=True before notifying.

The new patch is wired into the existing transformers_v5_compat() general plugin so it runs automatically in all vLLM processes.

^{Written by Cursor Bugbot for commit 1c192cc. This will update automatically on new commits. Configure here.}