Crunch dataset sources only read first file in directory

FileSystemDataset#getCrunchSource only uses the first path in the directory. It should either use all of them (if that's possible in Crunch), or it should use the directory (and fail if the dataset is partitioned - so you have to specify a leaf partition). The latter would be simpler, but doesn't currently work since Crunch's CompositePathIterable#FILTER doesn't ignore hidden files (only those that start with _).