Learning Locally, Communicating Globally: Reinforcement Learning of Multi-robot Task Allocation for Cooperative Transport