fix bug for Kruskal Wallis Test

[daikonradish]

fixes #148.

The Kruskal Wallis test requires there to be some underlying variation within the dataset. That means that the test should fail when provided with samples that are all of the same value.

This fix inserts a check for the first rank and the last rank within kruskalWallisRank. If the first rank is the same as the last rank, then there is no underlying variation, and Nothing is returned and propagated forward.

[Shimuuar]

kruskalWallis and kruskalWallisRank are part of public API, so this is breaking change. I'll try to remember what is Kruskal Wallis test and try to invent way to avoid breaking changes

P.S. Build problems with GHC 9.0 and 9.2 are

[Shimuuar]

P.P.S. Build problems with GHC 9.0 and 9.2 are due to some caching problem. Ignore them

[Shimuuar]

What is behavior of exiting tests: scipy throws exception,. R return NaN for p-value. So we need to answer what is desired behavior. Standard formula is not defined in that case but maybe there's sensible limit.

Numerical experimentation suggest that there's no sensible limit. If we take test for two large samples where almost all elements are identical except few. p-value depends only on number of elements that differ from mode and nearly independent of sample size. I think returning Nothing is correct choice.

I'd like to avoid breaking changes if possible

• kruskalWallisRank function is perfectly defined for samples where all element are same. No need to change anything
• Same for kruskalWallis. It returns NaN but it's quite reasonable way to communicate failure in numeric code.

P.S. Please rebase on top of master. I've fixed CI failures.