How Abstract Is Linguistic Generalization in Large Language Models? Experiments with Argument Structure