Adding PatchTransformerAggregation

[andyhuang-kumo]

Adding patch transformer architecture from dynamic graph learning literature as a PyG aggregation module.
Mostly based on "Simplifying Temporal Heterogeneous Network for Continuous-Time Link Prediction"

• added patch transformer aggregation module in torch_geometric/nn.aggr.patch_transformer.py
• added all related util functions in pt_utils.py
• passed precommit hooks
• added test_patch_transformer_aggr.py in `test/nn/aggr/ , test passed
• add documentation for helper functions
• integrate into workflow
• test on dataset

Notes:

• PatchTransformerAggregation requires edge features(edge_feats) and edge timestamps(edge_ts)
• PatchTransformerAggregation primarily focus on aggregating edge information for each node based on their past interactions, the embedding of edges are fed into the patch transformer blocks which learns interaction between these edges and by pooling over multiple patches, the representation of each node is generated. Lastly, this representation is merged with the input node features (x) to produce an output.
• PatchTransformerAggregation do not run self.to_dense_batch() at the moment thus do not reshape the input x tensor.
• the output dimension is set via the out_dim argument

[zechengz]

In general LGTM. Maybe @rusty1s can help to make another review when available.