Add statistical testing to compare single-cancer and pan-cancer results

[jjc2718]

A t-test with FDR correction would probably work fine as a first pass at this.

We should also compare signal to shuffled prediction.

[jjc2718]

This was somewhat addressed in #20, but we're currently using an unpaired t-test. It's likely that a paired t-test would have slightly better power, but we need to verify that the cross-validation folds being compared are actually the same between runs, in order to use the paired test.