Enable use of GAT architecture in Relational Deep Learning

[cpondoc]

As part of some exploration, we adapted the HeteroGraphSAGE layer in RelBench to create a model that uses Graph Attention Networks as opposed to GraphSAGE. Users can then choose to build their RDL pipelines using this GNN architecture for ablations on different datasets.

cc @JosephGuman @merceod

[rishabh-ranjan]

Hi, for this to be merged, can you please attach experimental results comparing the existing model with the GAT model you implemented. Thanks