On the Bottleneck of Graph Neural Networks and its Practical Implications