__array_ufunc__ for xarray

[shoyer]

Now that __array_ufunc__ has appeared in NumPy 1.13, we should be able to resolve a long-standing annoyance with xarray: the inability to use NumPy ufuncs like np.sin() directly on all xarray objects, such as xarray.Dataset.

I believe this could be straightforwardly accomplished by wrapping xarray.core.computation.apply_ufunc.

Some care should be taken to ensure that properly defer to other array types, as suggested in the NEP. This also could be a good time to resolve some of the xarray's broader consistency issues when interacting with and wrapping other array-like types.

[shoyer]

A related task would be finishing up this PR:

• Do not convert subclasses of ndarray unless required #1118: Do not convert subclasses of ndarray unless required

Ultimately, it would be nice to have some sort of registration system, especially for objects that aren't ndarray subclasses, that indicates that a class can safely be wrapped by xarray.

I'm thinking something like xarray.register_data_type(ArrayWithUnits) which indicates that .data for xarray objects can now be an ArrayWithUnits object. This would also possibly require passing a reference to a set of function to use on data containing these objects (e.g., for stacking or concatenating).

[shoyer]

I'm starting to work on this.