For a 1-feature linear model pred = w * x + b and the mean-squared-error loss against targets y, use autograd to compute the gradient of the loss with respect to w and b. Return a length-2 tensor [dL/dw, dL/db].