colors: ensure masked array data is an ndarray

[ngoldbaum]

This fixes compatibility for imshow plots with array data that is a unit-aware ndarray subclass, for example data from yt.units or astropy.units.

Take for example the following script:

import numpy as np
import matplotlib.pyplot as plt
from yt.units import km

arr = np.random.random((400, 400))*km

plt.imshow(a)
plt.show()

This produces the following traceback right now: https://gist.github.com/ngoldbaum/27effb41173859132fed08a0ad5485ee

This is not isolated to yt's units - if one tries the same thing with astropy's units the same issue arises. Pint does not have this issue because Pint's Quantity class is not an ndarray subclass, so when it is converted to a masked array in matplotlib.image, the units are stripped.

This function seems to expect result.data to be an instance of the base ndarray class. This patch just enforces that expectation explicitly.

[ngoldbaum]

Hmm... that failure on travis is worrying. Let me see if I can reproduce it...

[ngoldbaum]

Can't reproduce that failure locally...

[jenshnielsen]

Travis failures are transient and non related

[efiring]

Looks OK to me, but I suggest adding a brief comment so that someone doesn't look at it a year from now and "optimize" it away, thinking it is superfluous.

[tacaswell]

Still not have tagged a v1.5.2, should probably backport this?

[ngoldbaum]

If you can backport this for 1.5.2, that would be excellent :)

[tacaswell]

backported to 1.5.x as 8d9357e

[tacaswell]

Backported as it fixes a major Yt headache