Sobolev Acceleration and Statistical Optimality for Learning Elliptic Equations via Gradient Descent