Benchmarking Provably Safe Reinforcement Learning Approaches