Learning General Sparse Additive Models from Point Queries in High Dimensions