Handle evaluations of sparse-labeled test matrices differently

[thcrock]

When a test matrix is keyed on entity id and date, and have sparse labels, we want to be more deliberate about what the evaluations mean. We should:

• Sort and threshold the predictions with NaNs intact, and then remove the NaNs.
• Add to the evaluations table the number of labeled examples, number of labeled examples above threshold, and number of positive labels. This will help the researcher make sense of different metrics.
• Add a false positive rate metric