[Serve] Add on_request_completed hook and supports_rejection_protocol to custom request router API

[abrarsheikh]

• Add on_request_completed(replica_id, internal_request_id) lifecycle hook to RequestRouter, called when a request finishes (success or failure). This enables custom routers to perform cleanup such as releasing capacity tokens.
• Always register a done-callback on ReplicaResult (not just when autoscaling metrics are enabled) so the router is notified of every completion. Add a finally guard for the error path where the callback couldn't be registered.
• Add supports_rejection_protocol property so custom routers that guarantee capacity client-side can opt out of the rejection round-trip.
• Extract inline backoff logic into overridable _compute_backoff_s / _backoff helpers.

[gemini-code-assist]

Code Review

This pull request introduces several valuable enhancements to the custom request router API. The addition of the on_request_completed lifecycle hook, along with the robust implementation that ensures it's called even in failure scenarios, is a great improvement for custom routers needing cleanup logic. The new supports_rejection_protocol property provides useful flexibility for routers that manage capacity client-side. Extracting the backoff logic into _compute_backoff_s and _backoff helpers is a good design choice that improves extensibility.

The changes are well-implemented and accompanied by thorough tests. My review includes a few minor suggestions to add return type hints for better code clarity and type safety. Overall, this is a solid contribution.