WIP: Change LDOs to use masked products by default

[e-koch]

This supersedes the first attempts in #540.

Our issue stems from subclassing u.Quantity for all LDOs (Projection, Slice, OneDSpectrum) while trying to use the masking routines built for spectral-cube (LDO.filled_data[:], LDO._get_filled_data(np.NaN)).

Basically LDO.filled_data[:] should be used by default for LDO[slice], LDO.value, LDO.quantity, etc. as this matches the behaviour used for SpectralCube. But this interferes with how we currently have u.Quantity subclassed.

[e-koch]

@keflavich @astrofrog help would be appreciated here!

This may require disentangling u.Quantity as a subclass for LDOs, closer to what we use for cubes. A lot of the LDO code is already confusing/hacked together.

The current changes fix the primary issue of returning the data filled from the mask but fall apart when slicing an LDO as __new__ doesn't get called and most attributes don't get set.

[coveralls]

Coverage decreased (-86.0%) to 0.0% when pulling 920df34 on e-koch:fix_projection_fill_value into 1497c07 on radio-astro-tools:master.

[keflavich]

Questions from telecon:

• Can we remove Quantity as a parent class? What capabilities would we need to add to a base class / mixin class to minimize breakage?
• Alternatively, can we override .value to use the spectral-cube masking scheme?

If you try any of these approaches, keep notes here on what breaks.