Latent Safety-Constrained Policy Approach for Safe Offline Reinforcement Learning