A Bounded $p$-norm Approximation of Max-Convolution for Sub-Quadratic Bayesian Inference on Additive Factors