Towards Better Laplacian Representation in Reinforcement Learning with Generalized Graph Drawing