Gradient Inversion Attack on Graph Neural Networks