The Limits of Transfer Reinforcement Learning with Latent Low-rank Structure