The LogLinearChangelistClassifier class needs to keep hold of the weights dictionary so that it can get the score of individual features in order to filter things and pretty-print the individual scores. We should augment the UnnormalizedLoglinearModel class itself so that it can return that information directly.

One option is to update the FeatureValue class to return the score in addition to the raw value. (And for normalized LLMs, it should also have the probability.)

A different option is to leave the FeatureValue alone but add a LLM method which takes FeatureValues and returns their scores (and similar for probabilities).