Update object detection attacks for tfdsv4 label format

[lcadalzo]

The tfdsv4 pipeline outputs object detection labels with a batch dimension, e.g. for an image with N boxes, the new pipeline outputs scores with a shape of (1, N) as opposed to shape (N,). This will require minor modification to attacks in armory.art_experimental to account for this (as well as in metrics code?).

To yield the errors currently present on the tfdsv4 branch, simply run any carla object detection config. As part of this issue, we'll need to figure out if this impacts all object detection datasets or just certain ones. This issue can be considered fixed once we can run all object detection configs without error.