GraphAttentionConv may not deterministic

[MagicPF]

Thank you for providing such convenient and powerful framework, while in my practice, my project shown that when using your GAT layer in my model, the output is not deterministic, which make my model not reproducible, I have check other layers in TorchDrug like GraphConv, NeuralFingerprintConv，GeometricRelationalGraphConv and Torch_geometric's GATConv, these layers can maintain a deterministic performance once I fixed all the random seed. Also I found in TorchDrug's GAT implementation, the self.concat was never used during calculation, and compared with Torch_geometric's GAT, seems you performed an extra normalization on attention weights. I would like to consult whether I am correct.