Improving Generalization in Meta-RL with Imaginary Tasks from Latent Dynamics Mixture