unpack compressed files instead of streaming through them

We could drop the dependence on `GZip`, which is not actively maintained, by just calling `DataDeps.unpack` (which relies on the p7zip binary) to uncompress files. It would be easy to change the MNISTReader & co. to just do that.

This could solve issues like #118.

Additionally, we may want to go in the direction of saving processed versions of the datasets (e.g. a JLD2 save of the dataset object itself) for faster I/O.