RINGS integration for existing GNN pipelines

[jeremy-wayland]

Summary

• Introduces rings.integrations.SeparabilityStudy, a thin, framework-agnostic collector that slots into a user's existing PyG training loop instead of taking it over.
• Replaces the direction taken on graph-bench-wrapper (a GraphBenchEvaluator orchestrator with a train_func(dataset, evaluator) callback and a hard graphbench-lib dependency) — that shape was too invasive for users who already have full training pipelines.
• Adds a canonical PyG/MUTAG recipe in examples/pyg_separability.py and unit tests in tests/test_study.py.

Why

RINGS' actual integration surface is small: perturbations are already PyG BaseTransforms, and SeparabilityFunctor already accepts a plain Dict[str, np.array]. So instead of wrapping the training loop, this PR provides ~5 lines of glue the user adds around their existing code:

study = SeparabilityStudy(perturbations={...}, num_seeds=5)
for name, transform, seed in study.runs():
    dataset = study.apply(base_dataset, transform)
    score = my_existing_train_and_eval(dataset, seed=seed)
    study.record(name, score)
results = study.evaluate()

apply() uses isinstance against torch_geometric.data.Data/Dataset (replacing the brittle hasattr(..., "__len__") branching from the prior wrapper) and torch_geometric is imported lazily so import rings stays cheap.

[Pseudomanifold]

Looks very good to me :-)