Improve confusion matrix and score distributions

Right now, the confusion matrix and the score distribution plots rely entirely on the test set human scores. This means that if there are, say, 6s in the training data but none in the test data, the confusion matrix will only be 5x5 and the score distribution plot will only show 5 bars. Perhaps it would be better to take the union of the human scores from both the training and the test sets?

Locate the code that builds the confusion matrix and score distribution plots, then trace how human scores are collected from the training and test sets. Update the plots so their score categories use the union of both sets, and verify that scores present only in training still appear in both visualizations.

Written by the indexing model from the issue text.