kmeans: small fix with inf variance

[arnike]

Hi there!

There seems to be no check on the cluster size when it's decreased, yet it's used afterwards to update the variance. So, if the cluster size reaches zero, the variance is infinity. As a result, the cluster gets chosen in the next iteration and causes access to furthestPoint == data.n_cols, which is rather sad.
I simply added the same condition as in Precalculate.

[rcurtin]

Hi there,

Thanks for the contribution! I've merged it in, which will help reduce general sadness (to me, this seems like a good thing).

I also made a slight change: the way MaxVarianceNewCluster works is that it will find the cluster with maximum variance and pull points from it to use as the centroids of empty clusters. But it does this by precalculating (once each iteration) which cluster has maximum variance. However, the situation you've stumbled upon can only happen if EmptyCluster() is called enough times that every point is taken out of the maximum cluster. Therefore, I've modified things slightly in ec8a151 so that a new cluster with maximum variance is found if the maximum variance cluster has only one point in it.

[rcurtin]

One more thought -- would you like me to add your name/email to the contributors list? If so, let me know what name/email I should use. Thanks!