Slight inaccuracy in operation queues length metric

[daniel-savu]

Problem

As a result of #4689, a bug was introduced where we missed some spots where decrementing the queue metric should have happened. Most of those were caught in #4691, but the queues for some chains are still overestimating the length by ~1.4%.

Should understand why that is still happening. If all op_queue.pop(...)s have an associated op_queue.push(op) or an op.decrement_metric_if_exists() call the metric would be accurate, so it must be that some popped operations go out of scope without either.

The way I measured the inaccuracy was by checking the grafana metric against the number of items returned by the list_operations relayer endpoint.