Should we implement `np.median` in dask array, provided that it is along an axis that is unchunked? This would be similar to how we handle `da.linalg.fft`