Dense Backpropagation Improves Training for Sparse Mixture-of-Experts