ENH: Threshold mask in images_to_matrix

[cookpa]

Prevents cryptic error if the mask is not binarized, because the threshold was being applied to the numpy matrix but not the mask itself.

Also add some more tests

[cookpa]

Related to #641

@ncullen93 the matrix here was all zero because it was sampling a corner of the image, by adding the spacing arg I make the downsampled mask cover all of the image domain

s = [65]*img.dimension
mask2 = ants.from_numpy(np.random.randn(*s), spacing=[4.0, 4.0])
mask2 = mask2 > mask2.mean()
imgmat = ants.images_to_matrix(imglist, mask=mask2)

Still not sure what the Windows seg faults were about, but will revisit if they happen again

[ncullen93]

Great! My guess is the windows error was because there was a function within a function trying to access an image in the outer function environment.. perhaps something with the c++ references that is quit complicated. I removed that and it will hopefully be OK. Here is a simplified version:

def outer_fn(image, mask):
    def inner_fn(image):
        # mask not in "inner" environment but being accessed anyways
        return image + mask 
    return inner_fn(image)

[coveralls]

coverage: 81.145% (+0.004%) from 81.141%
when pulling 0796515 on image_to_matrix_mask
into 7d876d8 on master.