Learning Heuristics for Transit Network Design and Improvement with Deep Reinforcement Learning