OmniDocBench extension for caption to parent graphic relationships

[ethnhll]

While working on a reading order implementation, I came across a bug in the caption to parent graphic association algorithms in the base reading order model.

This is relevant because I noticed while testing and trying to verify the bug, that OmniDocBench keeps caption to parent graphic relationships in ground truth, but does not surface them as part of the benchmark itself.

I am wondering if there would be any interest and/or if folks think it would be useful to have a measure of caption to graphic relationship accuracy in this eval package? If so, I can spend some time refactoring the scripts I wrote up to do my own testing and present them as new benchmark additions.

Any thoughts?