Retry ack for claimed work items

[sylvesterkaczmarek]

Summary

Retry acknowledgment for the work item that was already claimed before polling for another item.

iter_work() and aiter_work() currently handle a transient work.ack() failure by backing off and then continuing the outer polling loop.

The current sequence is:

1. poll() claims work_1;
2. ack(work_1) fails transiently;
3. back off;
4. poll() again.

The original item is never acknowledged or yielded before the worker resumes polling.

Fix

Keep the claimed item in hand and retry its acknowledgment in an inner loop.

For sync and async pollers:

• transient failures retry ack() for the same work_id;
• no new poll() occurs between attempts;
• success resets the ack backoff and yields the original item;
• permanent 4xx failures retain the existing force-stop behavior;
• polling and stop semantics otherwise remain unchanged.

Regression coverage

Adds deterministic sync and async tests where the first ack returns a transient 500 and the second succeeds.

The tests require the operation order:

poll, ack, ack

and fail if a second poll() occurs before acknowledgment succeeds.