Learning a Distributed Hierarchical Locomotion Controller for Embodied Cooperation