Reduce memory usage when computing an array with missing data

[Cadair]

As of #301 we return a numpy array when data is missing, to prevent a situation where when you computed the array you got a dask array back.
However, the upshot of this is that dask appears to un-broadcast the array, so you require the full memory allocation of the array to compute an array with no data in it.
This is very inefficient.

We should see if there is a way to generate a dask array which uses much less memory for some or all NaNs.