Exploiting Low-Dimensional Structure in Bayesian Inverse Problems Governed by Ice Sheet Flow Models