Improve hourglass to handle different image resolution

[NathanMolinier]

Description

Images with greater resolution than 450 for each axis were not used for the training even if a 256x256 resizing was applied before being given to the hourglass network. Regarding the testing, a zero padding was used thus modifying the look and the effectiveness of the hourglass.

https://github.com/spinalcordtoolbox/disc-labeling-hourglass/blob/main/src/dlh/utils/data2array.py#L259-L272

This issue needs to be handled so the network may be used with different datasets.

[NathanMolinier]

In the PR #15 a choice was made to resize every input image using a 256x256 resolution, thus modifying the look of the images to improve generalisation.

Another solution could be to use cropping (center crop, crop based on the spinal cord centerline, ...) but some contexts information could be missing.

This difference will need to be investigated to keep the best solution.