A Provably Effective Method for Pruning Experts in Fine-tuned Sparse Mixture-of-Experts