Better support for parallelism & memmaping

[keflavich]

There are a few problems with how SpectralCube is set up for use with, e.g., joblib:

• SpectralCube.base should return the base of its _data
• __getitem__ should not use filled_data
• but, in order for __getitem__ to return meaningful data, we should have OneDSpectrum and Projection and Slice implement masking

@astrofrog - at this point, am I pretty much asking for nddata here? Can nddata do any of this stuff? Are we working in parallel with orthogonal to any other groups within the astropy collaboration?

If we get all this working, it will become easy to make some of the slicewise iterated things operate in parallel!

[keflavich]

On the base issue, it is needed to support this function

[astrofrog]

@keflavich - since we originally started SpectralCube, NDData has changed a lot (much more lightweight now) so it might be worth investigating whether we can start to merge back in that direction and solve some of these issues in combination with efforts there.

[keflavich]

@astrofrog finally looking at this again. So, should SpectralCube inherit from NDData and NDSlicingMixin? Presumably LowerDimensionalObject should, since that will save us writing new masking routines, but it looks like SpectralCube already implements nearly everything the base NDData and mixins provide, and SpectralCube does it differently for good reasons.