Though the ideal max_splits for
n < 1
million (or so) on x86 seems to be substantially
larger, enabling a roughly 15% speedup for such tests, this optimization
isn't general, and doesn't apply for
n > 1 million.
A too large max_splits can cause sort to take
more than twice as long, so it should be set on the low end of the
reasonable range, where it is right now.