Custom scoring of training samples for DTW.

[damellis]

Our current approach based on class likelihoods may not be the best for DTW. Maybe borrow the goodness score from MAGIC? This should probably be for both the confusion / overlap scores and the impact scores (which are basically duals anyway).

[damellis]

For each sample, we should probably at least show the average distance to all other samples in its class. This would give some sense of the "spread" of the class, plus the extent to which each sample is representative or not.