New mode to split prediction by starting tokens

We need a new way of splitting the text to create objects through attention maps. This time, the split will be done on NER starting tokens. This new Level could be called and exported as entity in the JSON files.

You'll need to update:

• https://gitlab.teklia.com/atr/dan/-/blob/main/dan/ocr/predict/attention.py?ref_type=heads#L47
• https://gitlab.teklia.com/atr/dan/-/blob/main/dan/ocr/predict/attention.py?ref_type=heads#L74
• and expose the parameter

You can test by making a prediction and viewing the resulting GIF via attention_map=True and visualizing the output JSON.