kvserver: better observability into slow latches

In https://github.com/cockroachlabs/support/issues/1952 we were looking at a number of traces that showed a request spending multiple seconds waiting for a latch.

It would be helpful to learn which overlapping latch was the slowest to release and, if possible, the batch holding it. In cases where the slowdown is workload-related (and not, e.g. network or disk slowness), this might provide valuable clues that allow the customer/TSE to self-diagnose without requiring expert assistance.

Jira issue: CRDB-22636