On-Robot Reinforcement Learning with Goal-Contrastive Rewards