Fix producer flush hangs when undrained batches expire

[zhechen-cs]

Completing an undrained expired batch with message ordering enabled removes a partition that was never muted. The resulting KeyError interrupts expiry processing, leaving other batches' futures unresolved and causing flush() to wait indefinitely.

Use discard() when unmuting partitions in both completion paths. Regression tests cover multiple undrained expired batches, late success/error responses after expiry, and normal successful completion with message ordering enabled or disabled.

Fixes #3159.

Validation:

• The three regression cases fail with KeyError on unmodified master and pass with the fix.
• pytest -q test --ignore=test/integration: 2743 passed, 1 skipped (Python 3.12.13).
• make lint and git diff --check passed.

Kafka broker integration tests were not run locally.

CI note: all 13 failing matrix jobs report test/net/backend/test_transport.py::TestKafkaTCPTransport::test_ssl_wrapper_transport raising ValueError: check_hostname requires server_hostname. The same test fails on clean base cd3f7938 and PR head 70d9c08e with Python 3.13.16 and the same development dependencies. This failure is independent of the producer expiry change.