Add a simplified version of KMKNN

[LTLA]

Effectively does kmeanspp only, holds onto the observation identities, and uses them directly as the centers. This saves us $\sqrt(n)$ extra calculations and cuts down on memory usage.

[LTLA]

Actually we can simplify this to (i) identifying the closest "seed" point in each cluster to its centroid, then (ii) replacing the centroids with that point, with some memory about what the identities of the points were, with possibly (iii) re-identification of the closest seed for each observation (see #2). Much of the rest of the search is the same as before.