Beyond Over-smoothing: Uncovering the Trainability Challenges in Deep Graph Neural Networks