Demystifying Oversmoothing in Attention-Based Graph Neural Networks