Clean up retry & long poll handling in retry client

There's a handful of ways calls can be considered long polls in the client code, by name or by (recently added) request extension. Make everything use extension instead of explicit name checking, and make sure metrics are set consistently for all long poll types. Probably also lift the code that always retries timeouts/cancels for task polling from the client into the worker poller so it's not special cased in the client.

Start by tracing the retry client’s long-poll detection and metrics handling, then inspect the worker poller’s task-polling retry behavior. Confirm all long-poll paths use request extensions and that timeout/cancel retries are handled by the worker poller rather than being special-cased in the client.