Tightening Bounds for Variational Inference by Revisiting Perturbation Theory