Collaborative Multi-Agent Deep Reinforcement Learning for STAR-RIS Assisted SWIPT: A Rate-Energy Pareto Frontier Analysis