MELD: Latent State Models for Meta-Reinforcement Learning