Semantic Parsing for Evaluating Large Language Models: Separating Linguistic Abilities with YARN