Learning Best Combination for Efficient N:M Sparsity