RefuteBench: Evaluating Refuting Instruction-Following for Large Language Models