Allow DataArray.to_series() without invoking sparse.COO.todense()

[khaeru]

This adds some code (from iiasa/ixmp#317) that allows DataArray.to_series() to be called without invoking sparse.COO.todense() when that is the backing data type.

I'm aware this needs some improvement to meet the standard of the existing codebase, so I hope I could ask for some guidance on how to address the following points (including whom to ask about them):

• Make the same improvement in {DataArray,Dataset}.to_dataframe().
• Possibly move the code out of dataarray.py to a more appropriate location (where?).
• Possibly check for sparse.COO explicitly instead of xarray.core.pycompat.sparse_array_type. Other SparseArray subclasses, e.g. DOK, may not have the same attributes.

Standard items:

• Tests added.
• Passes isort -rc . && black . && mypy . && flake8
  (Sort of: these wanted to modify 7 files beyond the one I touched; didn't commit these changes.)
• Fully documented, including whats-new.rst for all changes and api.rst for new API.

[dcherian]

Thanks @khaeru

Make the same improvement in {DataArray,Dataset}.to_dataframe().

I think only Dataset._to_dataframe is needed since DataArray.to_dataframe() calls that.

Possibly move the code out of dataarray.py to a more appropriate location (where?).

Perhaps to_series could be refactored to use _to_dataframe too?

Possibly check for sparse.COO explicitly instead of xarray.core.pycompat.sparse_array_type. Other SparseArray subclasses, e.g. DOK, may not have the same attributes.

👍 Should add something to the docstring once you figure this out.

This PR also needs tests.