State estimation for quaternion-valued neural networks with neutral delay