Learning to Execute: Efficiently Learning Universal Plan-Conditioned Policies in Robotics