Add bounding box to nutrition extraction

[alexgarel]

As we predict nutrition, we get the position of words that are part of nutrition.
From that and the OCR information, we can

While this is used then to crop nutrition at validation time, this is not exposed in the insight, so we can't provide a good visualization to users validation the insight.

1. compute the bounding box of the nutrition just after model prediction (while we have the OCR data)
2. add bounding box to nutrition prediction and insight data
3. let users propose a new bounding box as they validate the data
4. use information for croping nutrition image