Solving Linear Systems by a Neural Network Canonical Form of Efficient Gradient Descent