ProST: Progressive Sub-task Training for Pareto-Optimal Multi-agent Systems Using Small Language Models