smp-server: rework message expiration

[shumvgolove]

Problem: the expiration sweep iterated every non-empty queue, doing a per-queue probe and COMMIT to find that most had nothing to expire. Cost is proportional to the number of non-empty queues times the per-queue commit, while only a small fraction of queues hold expired messages.

Change:

• Messages-driven expiration: one indexed scan finds the queues that hold an expired message, then the unchanged delete_expired_msgs runs only on those, committing in batches.
• Adds idx_messages_expire (msg_ts, recipient_id) WHERE NOT msg_quota.
• Drops msg_queue_expire (column and idx_msg_queues_expire): no query reads it for a decision after this change, and removing it also removes the flag update that defeated HOT updates.
• Drops idx_messages_recipient_id_msg_ts (unused; covered by recipient_id_message_id). Keeps recipient_id_msg_quota (in use).

Verification:

• Functional: differential test (old schema vs migrated schema) over writes, quota markers, peek-ack, and expiration; msg_queues counters and messages are byte-identical.
• Work: on a 20,000-queue / 1,000-expiring dataset, delete_expired_msgs calls drop from 20,000 to 1,000 for the identical result, with discovery via Index Only Scan.
• Schema: server_schema.sql matches pg_dump 15 output; the down migration restores the prior schema except msg_queue_expire column order (added to the schema-test down-skip list, as with 20250320_short_links). Library and test suite compile.