Large Language Model-Based Reward Design for Deep Reinforcement Learning-Driven Autonomous Cyber Defense