Monitoring: improve reported ATProto subscribeRepos send delay

[snarfed]

We use the task_delay log-based metric to monitor whether we're behind on emitting records via subscribeRepos to the relay and other clients. We need to fix two things here:

• Right now, we include timestamps from clients reading historical events in the rollback window, which we shouldn't, since it makes the actual reported delay wrong
• This metric stops updating if we stop emitting events. We should regularly check and report the difference between our latest stored seq and the latest event we've emitted