Causally Correct Partial Models for Reinforcement Learning