Enhance Exploration in Safe Reinforcement Learning with Contrastive Representation Learning