Benchmarking safe exploration in deep reinforcement learning