Contrastive Variational Reinforcement Learning for Complex Observations