On the Trade-off between Over-smoothing and Over-squashing in Deep Graph Neural Networks