Replication delay metric

This implements a new replication delay metric. While we already measure how long a replication job is running with the latency metric, this only records the time from dequeueing of that job until it's finalization. What's not covered is the time it takes for a job to be processed after having been queued, which is what this new metric records. The total lifetime of a replication job is thus delay + latency.

Fixes #2530 (closed).