Training Differentially Private Graph Neural Networks with Random Walk Sampling