Human-compatible driving partners through data-regularized self-play reinforcement learning