correct tweak to entropy calculation

[isaaclegred]

This should be more robust, erroring if the weights are indeed zero in a transparent way, and also fixes what I think was the actual problem which was that a small number divided by a large number can be the zero float, so it's better to use an evaluation strategy of x*log(x) which is robust to zero values that may pop up

[reedessick]

I'm addressing this in #42 in a slightly different way (to avoid relying on scipy)