Safe Reinforcement Learning for Real-World Engine Control