ROSARL: Reward-Only Safe Reinforcement Learning