A Learning Theory Approach to a Computationally Efficient Parameter Selection for the Elastic Net