Broken attention mapping

With the introduction of the variables char_separators and tokens_separators the visualization of attention broke. These need to be integrated. I have created a branch that fixes these issues as well as fixes the regex that splits the text. These fixes can be found at https://gitlab.teklia.com/atr/dan/-/blob/fix-attention_vis/dan/ocr/predict/attention.py?ref_type=heads and https://gitlab.teklia.com/atr/dan/-/blob/fix-attention_vis/dan/ocr/predict/inference.py?ref_type=heads.