Towards Higher Pareto Frontier in Multilingual Machine Translation