Accelerating Generalized Linear Models by Trading off Computation for Uncertainty