A Modular Multi-Agent Reinforcement Learning Framework Guided by LLMs: Improving Quality and Efficiency of Treatment Planning in Lung Radiotherapy