kl_divergence and js_divergence do not propagate NaNs

[mauricelanghinrichs]

I noticed the following behaviour for the distribution-based KL and JS distances. In contrast to euclidean as a reference, kl_divergence and js_divergence distances do not propagate NaNs, leading to negative distances in the examples below. I don't know if this is desired behaviour; in my opinion they should rather propagate NaNs instead.

julia> using Distances
julia> kl_divergence([NaN, 0.1], [0.5, 0.5])
-0.16094379124341004

julia> js_divergence([NaN, 0.1], [0.5, 0.5])
-0.10051100363189411

julia> euclidean([NaN, 0.1], [0.5, 0.5])
NaN

[richardreeve]

@devmotion - #260 closes this as well as some other minor testing fixes. Could you check it and merge it if you're happy?