kv95 performance dropped significantly on Feb 18, as seen here: https://roachperf.crdb.dev/?filter=&view=kv95%2Fenc%3Dfalse%2Fnodes%3D1%2Fcpu%3D32&tab=gce

The initial drop was investigated in #62148, and was due to tracing. This was solved by #61777 (verified by cherry-picking onto the culprit PR #59992), but performance did not fully recover, so there were other performance regressions in the period following Feb 18. We need to figure out what these are and get back to the pre-Feb 18 baseline.