Efficient Model-Based Deep Reinforcement Learning with Variational State Tabulation