Make AFQDataset work with different input formats and make indexable

[richford]

Resolves #98
This PR

• makes AFQDataset initialization more general so that user can provide X, y, groups, feature_names, group_names, subjects, sessions, classes explicitly on init.
• create a new static method from_files() that takes all of the filenames that are in the current AFQDataset init method and returns an AFQDataset object with all of the data read in.
• Make AFQDataset more array-like by adding __getitem__ and __len__ methods and a shape parameter. This means users can do things like dataset[10:20] and also that AFQDatasets can be used as input to sklearn functions like train_test_split.
• Adds a lot of detail and doctests to AFQDataset's docstring

Note that I chose a different solution to dataset splitting than the one I proposed in #98. I think making the datasets indexable and therefore interoperable with scikit-learn's already existing, performant, and robust model selection routines is a much better solution than rolling our own split method.

This doesn't resolve the need for imputation methods, which we hint at in #98 but I now think that if we still want those, we should open up another issue and PR for that.

[arokem]

Overall, looks great! I think that a sphinx gallery example of using train_test_split with some real data should be helpful. But we can take that on a separate PR.