Show median instead of mean

[giordano]

The mean is a terrible estimator in the presence of outliers, in https://tealquaternion.camdvr.org/compare.html?start=dd0802b7462871a7d454af069fa0677c20d2edb0&end=0f2031fd150d453d8e86990e500f26c78c43adfa&stat=min-wall-time

I'd be more interested in seeing the median than the mean (or perhaps both, which is what BenchmarkTools.jl does).

CC @Zentrik.

[Zentrik]

Seems reasonable, I tend to ignore it anyways due to the overweighting of regressions (as they have infinite percentage whilst improvements are capped at 100%).

[giordano]

Thanks, this looks better now!

For the LLVM 21 PR, median and geometric mean both point to (slight) improvements, which is nice. The mean was pulled to the other side by a single large outlier.