Ecological Velocity Planning through Signalized Intersections: A Deep Reinforcement Learning Approach