Add support for bounding box labels

I'm handling support for image-formatted labels (i.e. pixel masks for segmentation), but we still need to support bounding boxes. Down the road it would be great to extend this to more complex polygons for things like Polygon-RNN, but that can wait. @avanetten, can you help implement the bbox stuff?

I'm not 100% familiar with the different bbox formats, how those records are generally stored, and how they're loaded and passed into nets. The places that I'm currently aware of that need greater bbox support are:

cw_nets.data.transforms Rotate and RandomScale classes: https://github.com/CosmiQ/cw-nets/blob/dev/cw_nets/data/transform.py. All of the classes built around the Albumentations DualTransform parent class need support for bounding box augmentations. The ones that we're importing directly from albumentations already have it.
cw_nets.data.datagen needs the same support - specifically in TorchDataset at the moment. FileDataGenerator (eventually to be renamed KerasDataGenerator) will also need it, but that's going to take some substantial restructuring first, so if #39 isn't resolved yet then don't worry about it for now.

More generally, since I'm not too familiar with what the inputs are generally like (are bboxes generally provided in individual files for each image? Are they just provided as coordinates in a pandas dataframe? etc.) it would be great if we could figure out a "standard" structure to expect users to convert them to. cw_nets could take that starting point and convert it to whatever is needed.

CosmiQ / cw-nets

Add support for bounding box labels #41