Iterators such as DynamicGraphTemporalSignal use edge_weights interchangeably with edge_attr

[lucianocheng]

Just a heads up that in DynamicGraphTemporalSignal and other iterators, edge_weights maps directly onto edge_attr into torch_geometric.data.Data.

Specifically, in dynamic_graph_temporal_signal.py, a Data object is passed back as a snapshot, where edge_weight (pulled from edge_weights) is passed into the edge_attr parameter of data.Data.

This has some key effects for dealing with snapshots in Pytorch Geometric Temporal:

1. When inspecting a Data object coming out of a DynamicGraphTemporalSignal iteration, edge_weights will appear in the edge_attr field.
2. DynamicGraphTemporalSignal accepts a kwargs parameter for snapshots where you can include arbitrary features with a snapshot. If you include a edge_attr field in this kwargs, it will overwrite your edge_weights into the Data object.
3. edge_weights can hold multiple features per edge, since edge_attr can support a multi-dimensional tensor per edge.

In the documentation, it's unclear in many places whether edge_weights can support multiple features (it can), or how edge_weights maps to edge_attr (it does). I propose we clean up the documentation to clarify this, or rename edge_weights as edge_attr in the code to be consistent with PyG / PyTorch Geometric.

In the meantime, highly recommend inspecting your snapshot torch_geometric.data.Data objects coming out of the iterators (such as DynamicGraphTemporalSignal to understand their shape before training.