Using large language models as judges with structured rubrics provides reliable evaluations for document extraction pipelines