Support for boolean/mask maps

[nwhitehorn]

It is often useful to have boolean sky maps (e.g. a mask) which are 64x larger than they strictly need to be if represented as doubles. This provides support for such a thing by providing a new sparse storage backend for maps based on std::vector, which stores entries as single bits. Semantics are the following:

• Comparison operators return a boolean map
• Multiplication of a map by a boolean map is implemented efficiently

To Do:

• Sparse backend for healpix masks
• Comparison operators as members of G3SkyMap (use boost::shared_ptr)
• Efficient mask multiplication operators for FlatSkyMap and HealpixSkyMap
• Tests
• Use masks everywhere
• Create masks from numpy arrays
• Convert mask to numpy array
• Support numpy ufunc interface for all() and friends, at least in the simple case
• Documentation

[arahlin]

Ran into a bug accessing the mask parent attribute:

In [1]: from spt3g import core, maps

In [2]: x = maps.FlatSkyMap(300, 300, core.G3Units.arcmin)

In [3]: m = x.to_mask()

In [4]: m.parent
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
Input In [4], in <module>
----> 1 m.parent

TypeError: No to_python (by-value) converter found for C++ type: boost::shared_ptr<G3SkyMap const>

[arahlin]

This PR is now feature-complete, up to the bug above, and an API converting masks to/from numpy arrays. It would be useful to add in the sparse backend handling for healpix maps with this PR, too, otherwise this won't do much in terms of memory efficiency for healpix maps.

[arahlin]

There are many instances of the numpy ufunc reduction wrappers (e.g. np.all()) being used for all, any or sum on sky maps or masks. These now fail because the corresponding C++ methods don't have the signature that the numpy ufunc machinery expects. I think the right solution is to add wrapper functions to skymapaddons.py that allow only the simple case (where the input arguments are axis=None, out=None) and otherwise raise a useful error. I might also consider adding similar functionality for mean and std, since these are also commonly used, so we can avoid many instances of producing accidentally dense maps. The alternative is removing the C++ functions and letting the numpy accessors make the maps dense in order to use its internal ufunc machinery, and I'd rather avoid that if possible. Thoughts?

[nwhitehorn]

I think the right solution is to add wrapper functions to skymapaddons.py that allow only the simple case (where the input arguments are axis=None, out=None) and otherwise raise a useful error.

This is the way. Anything else will be a lot of work and still be fragile.

[nwhitehorn]

Sasha's changes look good to me. I can't actually press "approve" because I apparently wrote too much of the code in the PR but I approved all the same.