Implement train method in KNN.

Currently, training of a knn is wrapped around the addNeighbour function.

1. Implement "train" for real
2. Implement some kind of clear function

Should train clear first, then add? Or just add? How to present these two modes in the API?