When filesort sorts a buffer and LIMIT is active, it first sorts all rows and
then discards the ones that don't fit into LIMIT. This sorts a lot of rows that
we know we're going to throw away anyway. Instead, find the right element using
nth_element first (if LIMIT is sufficiently smaller than the buffer size -- the
factor 2 has been chosen without much measurement), and then sort only the rows
we know we actually need. This speeds up the sorting phase of a simple string
sorting benchmark (as measured with EXPLAIN ANALYZE) by ~15%.

Change-Id: Id267c8e9ed264c32f7e338ff028efbe52fe9d428