Heterogeneous graph modelling and anlytics

[mjdhasan]

We want cuGraph to support heterogeneous graph analytics and modelling, because we are using cuGraph to run analytics and train graph models that includes heterogeneous nodes and edges.

Our current issues include

1. No native heterogeneous analytics in core cuGraph
   Core cuGraph lacks heterogeneous PageRank, centrality, or metapath analytics APIs, so classical analytics remain edge-type agnostic (cugraph-gnn/README.md).

2. Metapath-driven operations not first-class
   The samplers expose per-edge fanouts but do not provide metapath walk builders or metapath-aware neighborhood sampling, keeping our custom metapath simulators relevant (no metapath utilities present under python/cugraph-pyg/cugraph_pyg).

3. Dynamic graph updates still require reloads
   GraphStore invalidates and rebuilds the GPU graph whenever we insert edge batches; there is no streaming update API for evolving hetero graphs (see self.__clear_graph() in cugraph-gnn/python/cugraph-pyg/cugraph_pyg/data/graph_store.py).

4. Sampling constraints to watch
   Hetero loaders only accept COO layouts, directional subgraphs, forward-only temporal sampling, and lack temporal negative sampling support, so certain training recipes may need workarounds.

See this document for more context about the problem.