RxSafeBench: Identifying Medication Safety Issues of Large Language Models in Simulated Consultation