Population-aware Online Mirror Descent for Mean-Field Games by Deep Reinforcement Learning