Implement TimingStats better
I found some material on why what we're doing isn't great (the variance might end up completely wrong), and some examples that do it better on Wikipedia.
Good news is that the average times we measure are probably fine.
Edited by Dan Guest