Provide Variable initialization from np.full for Datasets

[itcarroll]

• Closes Creation of empty DataArrays in a Dataset #9468
• Tests added
• User visible changes (including notable bug fixes) are documented in whats-new.rst

Adds logic to variable.as_variable that handles tuples like (["x", "y"], 0) and (["x", "y"],) which results in a Variable with the given dimensions and a constant value (using np.nan if not given). The as_variable function takes an new optional argument called sizes that calling functions (notably core.merge) can construct from existing mappings. The construction of sizes added to core.merge is a possible performance concern.

I also considered putting the logic "higher up" in dataset.merge_data_and_coords, but that required an additional loop through the data_vars mapping and ended up with duplicate case checking in as_variable anyways.

Starting as draft for feedback and before updating any documentation. I've also discovered that the analogous enhancement to DataArray doesn't work with non-dimensional coordinates, which I'd like to take a look at for consistency.