feat(ml-api): split liveness/readiness probes and drain on shutdown

[pavsoss]

Summary

/health was a static {"status":"ok"}, so an orchestrator could route traffic to an instance that would fail every prediction, and SIGTERM dropped in-flight requests.

• GET /health/live — always 200 (process up).
• GET /health/ready — 200 only when the serving state is loaded and the spam-words DB and rate-limit store respond; otherwise 503 in the standard error envelope.
• GET /health retained as a backward-compatible alias.
• In-flight requests are tracked; on SIGTERM the app enters draining (readiness starts returning 503 so load balancers stop routing) and waits for in-flight requests to finish before exit.

Test plan

/health/live is 200 independent of dependencies; /health/ready is 503 when the model isn't loaded / a dependency is down and 200 when healthy; entering drain flips readiness to 503; /health behaves as before.

Closes #1009

[vercel]

@pavsoss is attempting to deploy a commit to the Aditya Sharma's projects Team on Vercel.

A member of the Team first needs to authorize it.