Feature/gnn

[silkspace]

Adds joint training GNN toy model and fits it on Cora dataset. For @tanmoyio and @webcoderz as templates for proj work.

[lmeyerov]

i'm unclear what problem this solves? rgcn for link prediction that handles both node+edge attrs and node+edge classes?

or for ER, how would this line up with a real dataset (hypergraph of ID -> attributes for raw input, and attribute<>attribute weighted equality scores of diff classes from feature engineering), + task (???)

[silkspace]

This gives them a base class that doesn't use DGL, is lightweight, and stubs the net while I am away.