Keeping track of a lower bound of stream ID where we've deleted everything below makes the queries much faster. Otherwise, every time we scan for rows to delete we'd re-scan across all the rows that have previously deleted (until the next table VACUUM).
