Crop should be more consistent in randomization

[BenjaminMidtvedt]

The default mode of Crop, "random", should crop each image in a list identically. However, this is achieved by seeding the rng. As a consequence, two images of size (N, N, 1) and (N, N, 2) will be cropped differently along the first dimensions.

I propose, instead, that we generate a random fraction for each dimension, which is subsequently parsed as an integer by multiplying by the dimension size. In this way, the randomization becomes independent of the size of each dimension and more consistent.

Moreover, we should allow crop and corner to be fewer elements than ndim, parsed as taking all elements and the corner at 0 for missing dimensions.