Change save format to npz from h5py

hdf5 may be overly complicated for our brain and model objects, and a higher potential for bugs. We should consider switching to npy/z files. Some more things to consider:

- Brain objects can be huge so saving with compression is probably a good thing
- We want to efficiently select and load subsets of the brain object so compression is probably not a good thing

Is there a solution that solves both of these problems? If not, what do we care about more?

The issue names no files, tests, or entry points. First compare the current h5py save/load path with npy/npz options, focusing on compression and subset loading; done requires a decided format and an agreed implementation scope.

Written by the indexing model from the issue text.