feat(queue): per-partition isolation and fix subscriber shutdown race conditions

[behinddwalls]

Summary

• Per-partition isolation in consumer: deliveries are dispatched to per-partition goroutines so a slow message on one partition does not block others. Per-partition ordering is preserved.
• Per-partition workers in MySQL subscriber: each leased partition gets its own polling goroutine, managed by a supervisor that reconciles workers based on current leases.
• Fix shutdown race conditions: separate workerWg tracks partition workers independently from the supervisor. workerWg.Wait() blocks before closing deliveryCh to prevent sends on a closed channel. Workers are removed from the map immediately on stop so reconciliation can start replacements without zombie entries.

Test plan

• TestConsumer_PerPartitionProcessing — verifies a slow partition A does not block partition B
• TestConsumer_PartitionOrdering — verifies messages within a single partition are processed in order
• TestConsumer_PartitionWorkerCleanup — verifies all partition goroutines exit cleanly on Stop()
• TestSubscriber_ReconcilePartitionWorkers — verifies workers are started/stopped based on lease changes
• TestSubscriber_PartitionWorkerPollAndDeliver — verifies a partition worker delivers messages
• TestSubscriber_StopAllWorkers — verifies all workers stop gracefully
• All existing tests pass via bazel test