Rethinking Model-based, Policy-based, and Value-based Reinforcement Learning via the Lens of Representation Complexity