Linear Representation Meta-Reinforcement Learning for Instant Adaptation