Metrics should be moved out of EndpointRegistry, and managed as a SubSystem

[eddyashton]

The per-endpoint metrics (number of calls, error, retries) has been a weird bridge between the Frontend and the EndpointRegistry for a while - primarily owned and updated by the Frontend, but available for the EndpointRegistry to call. It currently lives on the EndpointRegistry, meaning its implementation is visible to all apps, and they could even call increment_metrics_*. We have a better way to own + expose things like this now - as a subsystem. I think this should even be a private subsystem, as the only use case we have for exposing individual metrics is in the node_frontend.

[achamayou]

@eddyashton I think we need to revisit whether metrics are worth keeping or not, now that we are going to log per-endpoint. @lynshi what do you think?

[achamayou]

• Check endpoint logging is comprehensive enough, and if so, remove /api/metrics now that users have confirmed that it is unused.

[eddyashton]

Resolved by #6283.