performance-cpu-busy-waiting-when-vllm-server-is-idle-with-tensor-parallelism-e-0a5ca0dc

THEN Apply the fix from vllm PR #16226 to the file `vllm/distributed/device_communicators/shm_broadcast.py`. The patch replaces a busy-wait loop with a proper condition variable wait, eliminating unnecessary CPU consumption. For Docker deployments, exec into the container, edit the file directly, then restart the container.

In shm_broadcast.py, replace the busy-wait loop that checks a shared flag with a condition variable wait. For example, change `while not self._flag: pass` to `self._cond.wait_for(lambda: self._flag)` and ensure `self._cond.notify_all()` is called after setting the flag.