Add support for image with no annotations in RetinaNet

[fmassa]

🚀 Feature

RetinaNet currently doesn't support training with images that do not have object annotations.
This has been disabled due to potential issues with distributed training.

We should enable this back.

For more context on what issues might be present, see #2784 (comment)