Necessary and Sufficient Oracles: Toward a Computational Taxonomy For Reinforcement Learning