It would be great if the Iris cube (rectilinear) interpolation could run lazily, since this would enable a few of the ESMValCore preprocessor to work lazily as well.

Right now the cube data is immediately realised when setting up the interpolator:

and the _RegularGridInterpolator that does the actual weight calculation and interpolation does not seem to support lazy data due to the interaction with scipy's sparse matrices.

Could we maybe contribute in making the interpolation lazy by dropping the immediate data realisation, and by including support for Dask arrays in the _RegularGridInterpolator?