RefactorBench: Evaluating Stateful Reasoning in Language Agents Through Code