Make classification sample data into a dict and use tv_tensors
This MR replaces our custom BoundingBox implementation by tv_tensors.BoundingBoxes.
In the case of tbx11k, transforms were not being applied to bboxes. The data part of classification samples was turned into a dict, which now contains the image and the optional bounding boxes. Image samples are now tv_tensors.Image.
The changes make the classification more similar to the segmentation part.
Closes #85 (closed)
Edited by Daniel CARRON