Patch-level Routing in Mixture-of-Experts is Provably Sample-efficient for Convolutional Neural Networks