fix: prevent blocking scrapes while cache is cold

[dborysenko]

Summary

• On startup, prometheus_client's default Collector.describe() calls
  collect() to discover metric names. This meant REGISTRY.register()
  triggered a full DB collection synchronously, blocking the process for
  the entire query duration before the server could accept any requests.
• Once running, if a Prometheus scrape arrived before the background thread
  had finished its first collection cycle, collect() fell through to a
  synchronous fresh collection — blocking the scrape and racing the
  background thread, doubling DB load on every scrape until the cache
  was warm.

Three coordinated fixes:

• Override describe() to return [], bypassing the implicit collect()
  call during REGISTRY.register() and letting the server start immediately.
• In collect(), return empty immediately when the cache is enabled but not
  yet populated, instead of falling through to a synchronous collection.
• Move _start_background_collection() before REGISTRY.register() so the
  cache begins warming as early as possible, shrinking the cold window.

Test plan

• Tested on production server — metrics are still showing as previously
• No errors in container logs, collection time unchanged