Improving LLM-Generated Process Model Quality Through Reinforcement Learning: The Role of Reward Function Design