Return the four gradients (dW1, db1, dW2, db2) for this network:

• z1 = X @ W1 + b1, a1 = relu(z1), z2 = a1 @ W2 + b2
• the loss is mean softmax cross-entropy against labels y
• no autodiff