It is sometimes difficult to see which node lead to the creation or replacement of which new nodes in the graph. Storing provenance data to show this will help debug. Similar ideas are implemented in FX: https://docs.pytorch.org/docs/stable/torch.compiler_inductor_provenance.html
cc @titaiwangms @gramalingam @xadupre