Differentially Private Optimization with Sparse Gradients