fix(kv-cache): drain finite replay requests

[jay-tau]

Summary

• distinguish finite replay completion from hard cancellation
• drain all generated BurstGPT and ShareGPT requests before shutdown
• preserve non-draining behavior for duration, request-limit, autoscaler, failure, and explicit cancellation stops
• cover graceful and hard-stop paths with replay regression tests

Tests

• focused replay shutdown tests: 21 passed
• complete KV-cache test suite: 250 passed, 2 deselected
• git diff --check
• compileall

Fixes #839

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅