It seems it isn't always a valid kwarg. This bypasses that problem
by calling `np.array` directly (which is just what `np.copy` does
under the hood anyway).
This PR exposes NumPy 1.8+ padding functionality to all
users of scikit-image, regardless of their personal NumPy
version.
The improved (much better scaling with dimensionality)
version introduced in NumPy 1.8 is used.