cifar10 and mnist should be sharded

[ry]

cifar10/train is 150mb and mnist/train is 50mb - accessing a single element from either of these will trigger a download of the whole set.

Ideally we can split these datasets into ~ 2mb chunks so they can be downloaded progressively. If only one batch is inspected (like for debugging) only 2mb will be used.

I've already split cifar10/train images into these files and uploaded them (using this script)

http://ar.propelml.org/cifar10_train_images_00.npy
http://ar.propelml.org/cifar10_train_images_01.npy
...
http://ar.propelml.org/cifar10_train_images_49.npy