During the computation of benchmark scores, the train/test images are loaded from the assemblies file (predictions made by the user) using the train/test indices in the dataset documentation file. This isn't robust, as the order of images passed by the user might be different to the ones in the ground truth file.

This PR updates the code to use the ground truth training dataset file to obtain the paths of the training/test images to use for evaluation, and find those images in the assemblies file. If there are predictions for test images that are missing, a warning is raised a warning and evaluation is made as if no predictions are made for that image.

The same issue happened when computing OKS in DeepLabCut - missing predictions can lead to a shift in the indices. This has been fixed in deeplabcut/pose_estimation_tensorflow/lib/crossvalutils.py.