fix(seed): prevent duplicate seeding under multi-worker gunicorn

When the beta image runs with --workers > 1 (default: 2), every worker
calls lifespan() in parallel. Each worker sees an empty DB, then
seed_if_empty() runs concurrently, producing 2x (or Nx) demo entries.

Fix: wrap the seed operation in a process-wide FileLock. Even if a
worker loses the race it re-checks the table count inside the critical
section and returns without re-seeding.

Verified by cold-starting a 4-worker container: 8 items (was: 32).

requirements.txt

@@ -5,3 +5,4 @@ pydantic==2.9.2
 httpx==0.27.2
 python-multipart==0.0.10
 jinja2==3.1.4
+filelock==3.16.1