Provably Safe Reinforcement Learning: Conceptual Analysis, Survey, and Benchmarking