But How Does It Work in Theory? Linear SVM with Random Features