Bug: fix postgres backfill timeout

[SimmerV]

Summary

• Schema migrations moved to independent try block: The ALTER TABLE migrations (including from_node_id/to_node_id columns on mqtt_messages) were inside the same try block as schema.sql execution. If schema.sql failed for any reason, the migrations never ran, leaving the table without required columns and causing column "from_node_id" does not exist errors on every MQTT write.
• Backfill timeout fixes: The pool-level command_timeout=10 was overriding the SET statement_timeout = '120s' in the backfill method. Added explicit timeout=120 to all asyncpg calls in _backfill_mqtt_node_ids.
• Partial index for backfill performance: Added idx_mqtt_messages_backfill partial index on mqtt_messages(id) WHERE from_node_id IS NULL so the backfill query uses an index scan instead of seq-scanning millions of rows.