Improving code regarding data management, contrast selection, pre-processing and test computation

Description

This PR improves the code and solves several issues regarding niftii data management, multiple contrasts, text file edition and data preprocessing.

First, the file src/dlh/data_management/prepare_data_training_and_testing.py used to initialize the input data with pre-processing steps was deleted and now the data is loaded at the beginning of the training: this allows more flexible use of the hourglass --> only one script to call directly dataset with niftii images https://github.com/spinalcordtoolbox/disc-labeling-hourglass/issues/11
Multiple input resolutions are now allowed for the training https://github.com/spinalcordtoolbox/disc-labeling-hourglass/issues/14
Image orientation is now handled during the loading using the change_orientation method in Image class https://github.com/spinalcordtoolbox/disc-labeling-hourglass/issues/3
It is now possible to train the hourglass network on multiple contrast (for now T1w and T2w). Multiple contrast selection is still not possible for testing (for now only one by one) but it is now possible to specify the contrasts used for training before before running the tests. https://github.com/spinalcordtoolbox/disc-labeling-hourglass/issues/12