Dynamic Data Mixing Maximizes Instruction Tuning for Mixture-of-Experts