Avoid panic with almost-zero parameters for FisherF

[mstoeckl]

• Added a CHANGELOG.md entry

Summary

When m or n is a positive floating point value with 0.5 m = 0.0 (e.g. 1e-45 for f32, or 5e-324 for f64) the constructor of FisherF could panic in one of the unwrap() calls.

Motivation

This is one of the unusual-parameter panics that I found by fuzzing, and fixing it turned out to be simple. (There's only a few more simple panics from fuzzing, and I expect fixing them before tackling the much trickier Binomial and Hypergeometric samplers may make things easier. And I'm not in a hurry :-). )

Details

The workaround is to reject FisherF distributions with m or n almost zero instead of unwrapping. I expect that in practice people only use integral m and n.

[dhardy]

LGTM, thanks