GORDON: Graph-based Object-centric Rewards for Decomposition of Long-Horizon Manipulation