Add request-level tests for PartitionKeyResolver retry, cooldown, dedup, and cancellation

[m-szymon]

Current tests check synchronous bookkeeping (cooldown flags, in-progress slot membership), but not the request-level behavior of partition key discovery.

Include:

• transient DescribeTable failures retried with backoff, then cached on success
• permanent failures (resource not found / access denied) suppress further DescribeTable calls until the cooldown expires
• concurrent discovery for the same table issues exactly one DescribeTable call
• discovery cancelled mid-flight leaves no stuck in-progress marker

Expected outcome:

• retry/backoff and caching verified by counting actual DescribeTable calls
• cooldown window enforced at the request level
• dedup confirmed via real request counts, not just slot membership
• future discovery still works after a cancelled attempt

Part of #41.