Fix text splitting when plotting attention maps

Attention maps are not correct when --attention-map-level is "word" or "line".

This is because our regex match every character that is not \n (or for words), but we are currently using

text_split = separators.split(text)

instead of

text_split = separators.findall(text)

in split_text() (/dan/ocr/predict/attention.py)