nfm init should probably become a thin wrapper on datalad create

[effigies]

This was prompted by watching Yarik's talk in Miami. I think the reasons for this were:

1) datalad already wraps a lot of the git machinery we'd want 2) Annex and sub-datasets would be potentially useful for test data

The main downside I see is that datalad is quite heavy.