Towards Efficient Pareto Set Approximation via Mixture of Experts Based Model Fusion