Load image using torch + use training pre-processing function during prediction

As for the training, we should

1. Load the image using torch, as in https://gitlab.com/teklia/atr/dan/-/merge_requests/183/diffs#2e65ddf1b1c74ce181c3fddddced96b16c4af86d_240_232. So this part of the code will be removed (https://gitlab.com/teklia/atr/dan/-/blob/main/dan/predict/prediction.py#L81)
2. Add the Preprocessing argument to the command to use the same function as during training and call https://gitlab.com/teklia/atr/dan/-/blob/main/dan/transforms.py#L148