Cautious Adaptation For Reinforcement Learning in Safety-Critical Settings