Respect endpoint limits

Supervisor applies limits to endpoints that spend shared resources or protect sensitive flows. Limits can differ by route and deployment, so do not hardcode an assumed request count.

When a response includes Retry-After, treat its value as seconds and wait before sending the next request. Add random delay when many workers may retry together.

Retry reads after the delay. Before retrying a write, confirm whether the first request completed and use a stable record identifier when the endpoint supports one.