Memorize to Generalize: on the Necessity of Interpolation in High Dimensional Linear Regression