Benchmarking the Safety of Large Language Models for Robotic Health Attendant Control