Prometheus metrics for the machine-learning service

The monitoring docs note that only the server container currently uses IMMICH_TELEMETRY_INCLUDE — any chance the machine-learning service could get metrics too? /metrics on the ML container just 404s (tried on the current release image; there's no prometheus client in it at all).

It would be really useful. ML is the part that eats the GPU, and when thumbnails or facial recognition are running slow I currently have no way to tell whether it's inference latency, a pile-up of requests, or models being loaded. Request duration and counts, queue depth, maybe which models are currently loaded — that's the graph I'm missing.

If this is welcome I'd be happy to try a PR. Something like a MACHINE_LEARNING_METRICS setting (the ML config already uses the MACHINE_LEARNING_ prefix) that turns on a simple metrics endpoint.