Implement Learnable Structural Positional Encoding (LSPE)

[zsz00]

Since GNN lack of canonical positional information, it reduces the expressiveness of a GNN model and cannot reach ideal task performance. It is a must to support positional encoding for GNN model. Thus, it is a analog to Transformer model.

Learnable Structural Positional Encoding (LSPE) supports positional encoding learning from message-passing network.

Ref.
Graph Neural Networks with Learnable Structural and Positional Representations
https://github.com/vijaydwivedi75/gnn-lspe

[CarloLucibello]

This is known as RandomWalkPE in pytorch geometric https://pytorch-geometric.readthedocs.io/en/latest/modules/transforms.html#torch_geometric.transforms.AddRandomWalkPE

[CarloLucibello]

Closed in #273