Saving larger-than-memory objects to zarr using dask and xarray

[rabernat]

@ks905383

What ended up working was adapting their process, which seems to basically save a temp zarr store for every step of the problem (rechunking to one chunk in space, regridding, rechunking to one chunk in time)

In theory, Dask should be smart enough to stream all of these operations together without an intermediate write to disk. Maybe this would be a good workflow to share in Large GeoSpatial Benchmarks: First Pass.