Re-introduce a read-based compaction heuristic. LevelDB had a heuristic to trigger a compaction of a table if that table was being read frequently and there were tables from multiple levels involved in the read. This essentially reduced read-amplification in read heavy workloads. A read-based compaction is only selected if a size-based compaction is not needed.

In addition to reducing read-amplification, a read compaction would also squash deletion tombstones. So we'd get a scan performance benefit in not needing to skip the deletion tombstones.