Fix valid batch size to 1

We noticed that, for some experiments, the validation CER computed during training was really high, but when running the final evaluation of the best model, the validation CER was correct/as expected according to the training CER.

It appears that, when we do not resize the images, sometimes there can be a lot of padding inside the validation batches, which can lead to hallucinations in the padded zones, leading to very high CER. On the contrary, during evaluation, we use a batch size of 1, so no padding is applied, leading to a lower/correct CER value.

We should always use a validation batch size of size 1.