Add `datasets.map()` functionality to `ParaphraseDataset` class

[jscuds]

Would be much faster than current implementation of processing each sentence pair individually.