Safe Reinforcement Learning as Wasserstein Variational Inference: Formal Methods for Interpretability