fix(api): avoid worker-hanging db reuse

[risu729]

Summary

• stop caching the Postgres/Drizzle client in Worker global scope
• create request-local Postgres.js clients with SSL, disabled type introspection, and short idle lifetime
• keep realtime WebSocket auth from performing a DB user upsert before the upgrade

Why

Live tail logs still showed 1101 cancellations on /api/realtime, and live curls showed the same cancellation on DB-backed public routes like /api/pois. Cloudflare's runtime can cancel requests when a Promise is tied to global or stale request-context I/O and can never settle. The previous DB singleton kept the Postgres client in module scope across Worker requests; realtime auth also hit that DB path before upgrading the socket.

Verification

• bun run check
• live observed before fix: GET /api/pois returns Cloudflare 1101 on current deployment