Optimal Transport Perturbations for Safe Reinforcement Learning with Robustness Guarantees