fix(ml): more graceful shutdown

[mertalev]

Description

This PR addresses some of the issues with the ML service's automatic restart.

1. The service now performs more checks before beginning a shutdown.
   • In particular, it now checks that there are no active requests and that no model is being loaded/downloaded (i.e. the model loading lock is not acquired).
2. The service does more cleanup.
   • It now shuts down the thread pool and closes log handlers.
   • It lets Uvicorn perform its own cleanup by sending a shutdown signal instead of canceling async tasks and stopping the loop.
3. Because of the use of a shutdown signal, it presents an INFO log instead of an ERROR log when restarting.
4. It sets the LD_BIND_NOW flag, which results in more deterministic RAM usage.

How Has This Been Tested?

All tasks work as normal and the server is able to restart after idling. RAM usage as inspected from htop is normal as well, including after restart.

[cloudflare-pages]

Deploying with    Cloudflare Pages

Latest commit:	c44915b
Status:	✅  Deploy successful!
Preview URL:	https://01c68da6.immich.pages.dev
Branch Preview URL:	https://fix-ml-improve-shutdown.immich.pages.dev

View logs

[jrasm91]

Nice!