Preserve fetched original edges for hetero PPR

[mkolodner-sc]

Summary

This PR adds opt-in preservation of original graph edges for heterogeneous PPR sampling.

When include_original_edges_in_ppr_subgraph=True, the PPR sampler now emits original graph edges whose endpoints are already present in the PPR-selected node set and whose adjacency rows were fetched during PPR traversal. The default remains False, which is the more faithful behavior relative to PyG’s get_ppr API: output only virtual seed-to-PPR-node edges.

Changes

• Stores fetched PPR adjacency in the C++ PPRForwardPush cache with optional edge IDs.
• Adds C++ extraction for original edges from completed PPR cache state.
• Reuses the same extraction path for typed and untyped heterogeneous PPR.
• Deduplicates shared fetched rows across typed PPR channel states.
• Keeps Python throughput-friendly:
  • no extra _sample_one_hop pass
  • extraction runs via run_in_executor
  • pybind wrapper releases the GIL during C++ cache scanning
• Preserves edge IDs when available so existing edge-feature fetching can continue to work.
• Leaves homogeneous Data support explicitly unsupported for this flag.
• Does not include GraphTransformer/model consumption changes.

Output Behavior

For heterogeneous PPR:

• Virtual PPR edges remain under ("seed_type", "ppr", "neighbor_type").
• Preserved original graph edges are emitted through the regular sampled-edge channel.
• Original edges follow the same edge_dir orientation convention as k-hop sampling.
• The preserved original-edge set is not guaranteed to be a full induced subgraph because it is limited to adjacency rows already fetched during PPR.

Validation

• make -C gigl-core unit_test_cpp
• Focused distributed sampler tests:
  • test_typed_ppr_sampler_loader_outputs_channel_attrs
  • test_ppr_sampler_can_preserve_original_edges_for_selected_nodes
• git diff --check
• make -C gigl-core check_format_cpp
• ruff check on touched Python/stub files

Local Benchmark

Synthetic no-edge-feature dataloader benchmark, flag off vs on:

Degree	Mean Delta	Mean % Delta	Original Edges / Batch
4	-0.014 ms/batch	-0.13%	~10.5k
8	+1.555 ms/batch	+12.83%	~33.3k
16	+2.458 ms/batch	+14.12%	~77.8k

Average across these cases: approximately +1.33 ms/batch mean overhead. While nonzero, this slowdown is relatively negligible.