Ability to score pipeline against ground truth test data.

[damellis]

Relies on #144.

[damellis]

This likely requires multiple evaluation metrics. We probably want to supply a few of our own and allow the example author to write their own. For instance, for static pose recognition, this might be about the percentage of the time in which the predictions match the ground truth. For event-based prediction (e.g. gesture recognition or beat detection) it might be about the total (or mean) delta between the ground truth event times and the predicted event times. Etc.