Use dask instead of numpy masked arrays

[pnuu]

Currently PySpectral uses masked arrays from Numpy to handle the data arrays. For better integration and SatPy support, these should be converted to dask arrays.

[djhoese]

As mentioned on slack, there has to be a decision made about this. Does pyspectral add a hard requirement on dask? Most numpy operations should work with dask, at least the arithmetic and ufunc ones. If pyspectral can be updated and assume that NaN is used as a fill value instead of depending on masks then it should be compatible with both in most cases.

[pnuu]

Yes, I agree. I noticed this after starting to convert the unit tests to the daskified version, and found out that there are soooo many use-cases that don't work with my dask conversion. SatPy was happy, but I need to think more if this is feasible for other (non-SatPy) use(r)s.

[pnuu]

Running a full disk scene of SEVIRI with SatPy to create day_microphysics composite takes 7.4 seconds, and with the fully daskified version 4.8 s. I think we need to find a way where all the use cases can still work.

[pnuu]

OK, found the way to do it: only daskify the NIR reflectance part, and adjust the documentation so that it doesn't have any references to masked arrays. If someone has used the reflectance computations without SatPy, there's a need to adjust the calling script to a) to do .compute() to get the result, and b) use np.isnan(arr) instead of arr.mask to get the invalid values.