Fix excessive regrid memory use

[davidhassell]

Fixes #623

Notes

• The previous approach loaded both the source and destination grid masks onto the weights, which had to be a dense numpy array since scipy sparse arrays do not support masks.
• The new approach keeps the weights in a scipy sparse array, but creates a separate boolean mask to applied to the regridded data.
• The change to cf/regrid/regrid.py pre-creates the sparse matrix representation in a returned regrid operator. This makes sense because the purpose of returning the operator is for efficiency in repeated future uses, so it makes sense not to have to do this every time.
• The change from coo_array to csr_array is because the latter is much more efficient at row slicing.

[sadielbartholomew]

FYI I just added one minimal commit to resolve the lone trivial merge conflict (new Changelog entries to combine) so I can review without it being in the way.

[sadielbartholomew]

Clever approach which I agree with as covered in the notes of the opening comment. The implementation is sound and doesn't call for any new testing. The performance improves very significantly in test cases I played around with and did basic profiling on, so passes with flying colours in that respect.

Therefore overall this is all good, though I've raised some minor comments and a question. Please consider those and then merge when ready. Thanks.