Allow for validation split as attribute of every Data object.

[jgallowa07]

The way in which we perform cross validation now, is quite slow as we must; convert mutations to be with respect to the reference sequence, build the one-hot matrix, and finally compile the predictive function before making predictions. One thing that would drastically speed things up (and simplify the interface / source code) would be to add the option to split the full dataset at the time of Data initialization (the place to do that would be here). Then, getting validation loss would be as simple and fast as getting loss - and then it would be trivial to get conditional loss.