A Psycholinguistic Evaluation of Language Models' Sensitivity to Argument Roles