Add from_relbench utility to convert RelBench databases to HeteroData

[AJamal27891]

Description

This PR is Part 1 of 4 in splitting the monolithic Warehouse Intelligence system (#10353) into modular, composable pieces, as requested by the core maintainers.

This PR introduces the from_relbench utility to torch_geometric.utils.relbench. It allows users to convert complex, multi-table databases from the RelBench (Relational Deep Learning Benchmark) environment directly into PyG's native HeteroData format.

Addressing Assessor Feedback:
This specifically addresses @wsad1's feedback from #10353: "why do we need a RelBenchDataset?"
Based on that guidance, I have completely dropped the custom RelBenchDataset class wrapper. Instead, to align with PyG's stateless data processing philosophy and avoid reinventing the wheel, this PR only introduces a pure utility function. It decouples PyG's dataset classes from RelBench's internal state.

Proposed Changes

• Added torch_geometric/utils/relbench.py housing the from_relbench conversion utility.
• Added exhaustive unit tests in test/utils/test_relbench.py utilizing dummy fallback flags to ensure rapid CI execution without massive database downloads.
• Updated CHANGELOG.md.

(Note: Parts 2, 3, and 4—covering the Warehouse Transforms, SAGEConv multi-task models, and LLM G-Retriever integrations—will follow in subsequent linked PRs once this foundational data layer is approved.)

---

Ref: #10353
Partially Closes #9839

[codecov]

Codecov Report

❌ Patch coverage is 97.77778% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 84.21%. Comparing base (c211214) to head (9aa7a26).
⚠️ Report is 185 commits behind head on master.

Files with missing lines	Patch %	Lines
torch_geometric/utils/relbench.py	97.72%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master   #10628      +/-   ##
==========================================
- Coverage   86.11%   84.21%   -1.91%     
==========================================
  Files         496      511      +15     
  Lines       33655    36058    +2403     
==========================================
+ Hits        28981    30365    +1384     
- Misses       4674     5693    +1019     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[puririshi98]

can you add an example of training a GNN+LLM system on this based on the code from examples/llm/txt2kg_rag.py

[AJamal27891]

can you add an example of training a GNN+LLM system on this based on the code from examples/llm/txt2kg_rag.py

This PR includes examples/relbench_example.py — a lightweight hetero GNN example that demonstrates from_relbench end-to-end:
from_relbench → HeteroData → SAGEConv + to_hetero() → node-level regression (championship points, MAE 9.2 → 2.1 over 30 epochs, <30s on CPU).
Adding GRetriever end-to-end here would require bridging heterogeneous → homogeneous graphs (from_relbench produces HeteroData with multiple node/edge types, but GRetriever expects homogeneous input (x, edge_index, batch)). That bridging involves projecting all node types to a common dimension, concatenating, and remapping edges — model-level code that belongs in a dedicated PR, not the data utility PR.
The full GNN+LLM integration with GRetriever (using to_homogeneous) will land in a follow-up PR as part of the split from #10353:

PR	Scope	Status
#10628 (this)	from_relbench utility + tests + hetero GNN example	✅ Ready
PR 2	to_hetero_edges bridging utility	Planned
PR 3	GNN+LLM models + GRetriever example (w/ to_homogeneous)	Planned
PR 4	End-to-end RAG pipeline	Planned

The end-to-end code already exists in #10353 this split keeps each PR focused and independently reviewable suggested by @wsad1 .

[AJamal27891]

CI update: The pytest failures were caused by tabulate 0.10.0 (unrelated to this PR). Resolved in #10634 by @rusty1s. Branch rebased on latest master.

[puririshi98]

lgtm can you just share a log of running the example to convergence

[AJamal27891]

lgtm can you just share a log of running the example to convergence

Loading RelBench rel-f1 dataset...
Done in 0.26 seconds.
Graph: 9 node types, 26 edge types

Training 30 epochs on "standings" point prediction...
Target stats (train): mean=6.25, std=13.49

Epoch: 001, Loss: 0.8340, Train MAE: 7.29, Val MAE: 7.36, Test MAE: 7.31 points
Epoch: 005, Loss: 0.2883, Train MAE: 5.02, Val MAE: 4.90, Test MAE: 4.89 points
Epoch: 010, Loss: 0.1753, Train MAE: 2.94, Val MAE: 2.91, Test MAE: 2.90 points
Epoch: 015, Loss: 0.1473, Train MAE: 2.78, Val MAE: 2.76, Test MAE: 2.72 points
Epoch: 020, Loss: 0.1134, Train MAE: 2.56, Val MAE: 2.50, Test MAE: 2.49 points
Epoch: 025, Loss: 0.1037, Train MAE: 2.44, Val MAE: 2.37, Test MAE: 2.37 points
Epoch: 030, Loss: 0.0890, Train MAE: 2.22, Val MAE: 2.19, Test MAE: 2.18 points

Final — Train MAE: 2.22, Val MAE: 2.19, Test MAE: 2.18 points

[puririshi98]

lgtm, @akihironitta @wsad1 to merge