Improving analysis of varation of experiments in math.variation (expected value + confidence interval)

[gfursin]

We currently have a simple way of analyzing variation of experimental results in math.variation in ck-analytics repo. We use density via gaussian_kde and can detect multiple expected values, i.e. to warn that there are several states and features to separate them should be found. However, we do not calculate confidence interval, etc (since which such approach we do not need it, but instead we look at the delta and if there is more than 1 expected value).

We should check that it works correctly and possibly improve it (calculate confidence interval, etc)...

Moved from CK: mlcommons/ck#26