Safe Reinforcement Learning From Pixels Using a Stochastic Latent Representation