[krishan711]

Description

Two improvements to the message queues, prompted by how the SQL queue behaves in production:

Graceful shutdown

• New abstract MessageQueue.release_message: hands a claimed message back to be claimed again straight away without using an attempt (SQL and Cosmos keep postCount; SQS and AQS make it visible again).
• MessageQueueProcessor.run/run_batches now handle SIGTERM/SIGINT (and request_shutdown()): the message being processed finishes, claimed messages not yet started are released, and the idle sleep stops immediately. A second signal falls back to the default handler.
• A message whose processing task is cancelled is released instead of waiting for its lease to expire (which also used up an attempt).
• Previously every deploy killed the in-flight job, which then sat for 300s until its lease expired and counted as a failed attempt.

Housekeeping

• New abstract MessageQueue.delete_expired_messages: SqlMessageQueue deletes succeeded/deduplicated rows older than retentionSeconds; other queues remove messages on completion and return 0.
• The processor calls it every deleteExpiredMessagesIntervalSeconds (default 600), logging rather than raising on failure.
• SqlMessageQueue.get_message_count no longer deletes rows; shouldSkipRemovingNonRetained is removed (breaking).

Also adds shouldProcessInParallel to run_batches so consumers can try processing a claimed batch concurrently.

Tests: release on every backend (including after a lost lease), shutdown finishing the current message and releasing the rest, a real SIGTERM stopping an idle run, a cancelled message being released without an attempt, and the processor deleting expired rows. 141 queue/lock/database tests pass locally against sqlite, Postgres, Cosmos, ElasticMQ and Azurite.

Note: make type-check on main currently reports 35 Unpack errors in core/store/*; they are identical without this change.

Checklist:

• I have updated the CHANGELOG with a summary of my changes