ReRoGCRL: Representation-based Robustness in Goal-Conditioned Reinforcement Learning