kvserver: investigate mvcc gc queue overload

I a recent experiment, I forced all replicas to be enqueued to the MVCC GC queue every 10 minutes. I ran sysbench write-only, and noticed that the performance took about 10% hit ~10 minutes after the workload started, probably because of MVCC GC.

I think we could investigate what happens during the MVCC GC, and if we can potentially come up with some potential optimizations. I think there are 2 things that happen: (1) MVCC GC queue + scanning and deleting the data, and (2) pebble compactions that happen after that. I think both (1) and (2) combined caused the performance to take the 10% hit.

Jira issue: CRDB-56056