Allow converting an image to a numpy array to raise errors

[radarhere]

Resolves #5084

The following code runs without raising an error.

from PIL import Image
import numpy as np
im = Image.open("Tests/images/truncated_jpeg.jpg")
np.array(im)

Even though an OSError is raised if __array_interface__ is accessed directly.

from PIL import Image
im = Image.open("Tests/images/truncated_jpeg.jpg")
im.__array_interface__

This is a bug in numpy.

This PR works around that bug by replacing __array_interface__ with __array__.

[hugovk]

Thanks for chasing up. Let's open an issue at NumPy as well?

https://gitter.im/numpy/numpy?at=6067b4ef3153ce63a3b6dcad says:

Its a bug, but I couldn't just rip the bandaid off it because of shapely. Should do that, but a PR to do a deprecation would be welcome also of course ;)

Maybe they won't have it deprecated and fixed by the next Pillow release in 3 months but let's at least give them a chance, and review this before release. I've added it to the 8.3.0 milestone.

[radarhere]

Ok, raised at numpy/numpy#18723

[radarhere]

NumPy 1.21 will deprecate the error catching, raising a warning in this situation.

It is expected that NumPy 1.23, a year later, would start to raise an error.

[hugovk]

Thanks for following up!

[radarhere]

numpy/numpy#20835 has been merged into NumPy, so NumPy 1.23.should raise an error.

[radarhere]

NumPy 1.23 has been released, so I've created #6394 to revert this.