Fixes for complex numbers

[shoyer]

Fixes #553

Also ensures we skip NaN when aggregating with complex dtypes.

Still needs release notes.

[davidtrem]

This is great ! Thanks!
One question still: what is the point of applying .real or .imag to dataset ?
As far as I understand xray.Dataset might be inhomogeneous (a mix of non homogeneous datatypes).
It doe make sense for xray.Dataarray though.
Maybe I was misleading by reporting incorrectly a pandas dataframe bug pandas-dev/pandas#10921 (comment) ?

[shoyer]

One question still: what is the point of applying .real or .imag to dataset ? As far as I understand xray.Dataset might be inhomogeneous (a mix of non homogeneous datatypes).
It doe make sense for xray.Dataarray though.

If every data variable in your Dataset has numeric values, .real and .imag are perfectly well defined. If not, you'll get an error message when you try to use them. This is similar to the case for other operations that only work on numbers (e.g., binary arithmetic).

My two cents is that .real and .imag would also be appropriate for pandas.DataFrame, even though they only make sense for certain dtypes.

[davidtrem]

OK. Thanks for clarifying 👍