Support axis/axes/keepdims in apply_gufunc()

[shoyer]

As of NumPy 1.15, NumPy's generalized ufuncs now support axis/axes/keepdims arguments:
https://docs.scipy.org/doc/numpy/release.html#generalized-ufuncs-now-accept-axes-axis-and-keepdims-arguments

It would be nice to support these in dask.array.apply_gufunc(), too. This would be useful for implementing functions like median (#3819).

[mrocklin]

cc @magonser

[shoyer]

One design decision to consider is whether axis/axes should be passed on to the wrapped functions (which might require numpy 1.15), or if we should simply do the axis reordering/insertion on the dask side.

I would lean towards the later since it isn't much more difficult, and imposes fewer requirements on the wrapped function. It's similar to what we currently do in xarray.apply_ufunc().

[magonser]

I'd be happy to get my hands on it - though won't get to it before Sep.