RD2: Reward Decomposition with Representation Disentanglement