Backpropagate MSE loss through a one-hidden-layer NumPy network to compute gradients for every weight and bias.
