perf: bulk materialization publication writes

[suguanYang]

Summary

• replace ORM row-by-row map-unit/token persistence with SQLAlchemy Core bulk batches
• cap demo materialization database publication with a cross-instance Redis leased semaphore (default concurrency 2)
• add Logfire stage timing for uploads, base rows, publication, snapshot, commit, and semaphore wait

Evidence

• local production dump: 50k token rows, ORM flush 12.54s vs Core bulk 5.74s (~2.2x faster) with the existing five indexes
• Core bulk wall time: concurrency 1 = 4.63s, 2 = 4.99s, 4 = 6.85s
• no production schema, index, or data changes

Validation

• demo contract: 10 passed
• retrieval contract: 16 passed
• worker contract: 330 passed
• Ruff and Pyright passed