running time.
Probability of getting lots of bad splits is very small. With
high probability we divide the list into fractional pieces (instead of
just nibbling off a constant-size piece). This is enough balance to
get asymptotic running time.
We could also prove an lower bound on the expected
running time (but we won't).
