benchmarks without stats because the samples are not normally distributed

[fonsp]

Sort-of reverting #52 because our samples are not normally distributed, the calculated stddev is super high.

This PR tests the ratio between medians with a fixed rate (20%) again.

But! In #52 I added some nice things:

• Tuning the benchmark before running it (this reduces spread)
• Using the tuned parameters from one benchmark for the other, for a fair comparison.
• Learned that median is more stable against outliers than mean

So this should still be more stable than before #52!