Clean the samples passed to the training function

A lot of unused parameters are sent to the training code, generated during the collate function (https://gitlab.com/teklia/atr/dan/-/blob/66f0f92db68d23c1312c617c8556169b5423e18c/dan/manager/ocr.py#L547). We should clean this so only pass the used parameters (labels, images, names, and ids at least).