Sub-optimality of the Naive Mean Field approximation for proportional high-dimensional Linear Regression