[AI models] Pick patches

[normanrz]

To train new AI models, we could have a tool that automatically generates bboxes to be annotated. The size would match our model fov. We could make sure that the boxes are not sampled in useless regions (e.g. all black). The number of boxes should be user-definable but with a useful default.

This feature is inspired by a feature in Empanada.