Fractional Gradient Descent with Matrix Stepsizes for Non-Convex Optimization