Scaling Law for Stochastic Gradient Descent in Quadratically Parameterized Linear Regression