Automated Evaluation of Retrieval-Augmented Language Models with Task-Specific Exam Generation