Restricting to the chip architecture maintains the quantum neural network accuracy, if the parameterization is a $2$-design