Currently all of the sorted data is spilled in one burst of work, which can overwhelm the system and cause replication lag spikes that interfere with the user's foreground workload. We should instead incrementally spill the sorter data over the course of the collection scan.