Feature/multiclass clf

This adds the basic support for multiclass training.

This does produce end-to-end results, but it shirks things like visualizing the training batches and computing multiclass metrics, for which there is currently not support.

What it does do it:

• return a probability of each class for every image given
• at train time the network is encouraged to predict high probabilities for all true classes, not just a single one.

What needs to be done:

• The output of the predictor will likely need to be munged to return a Multiclassification object or extend the existing Classification object such that it understands that there can be multiple true / pred values.