On the Robustness of Safe Reinforcement Learning under Observational Perturbations