K-means++ Style initialization

[drroe]

Currently K-Means initialization is more or less random, followed by picking points distant from each other. There should be an option to use points of high density for initialization (similar to DBscan), aka Kmeans++ https://en.wikipedia.org/wiki/K-means%2B%2B