Parse the tokens

We should parse the tokens in https://gitlab.teklia.com/atr/dan/-/blob/3eda9edb4931ff5c0dd05369ef3ee375fd5cc38b/dan/ocr/predict/prediction.py#L448 instead of https://gitlab.teklia.com/atr/dan/-/blob/3eda9edb4931ff5c0dd05369ef3ee375fd5cc38b/dan/ocr/predict/prediction.py#L331 to parse the file only once.