Score Models for Offline Goal-Conditioned Reinforcement Learning