[Elastic][Distributed] Add configurable randomized backoff after failed rendezvous state writes (#185930)

At large scale, concurrent rendezvous updates can cause repeated CAS conflicts. The executor currently retries without backoff, potentially increasing backend load and delaying rendezvous progress during elastic restarts.

This patch adds a small randomized backoff (0–0.3s) when the existing sync() call reports a failed write. It preserves the current synchronization point and state-machine structure.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/185930
Approved by: https://github.com/d4l3k