Adapt fidelity and unfaithfulness to HeteroExplanation

[Mellandd]

Adapts the explainability metrics in torch_geometric.explain.metric to heterogeneous graphs (#9112 ).

fidelity and unfaithfulness previously assumed a homogeneous Explanation and accessed explanation.x / explanation.edge_index with plain tensor masks. When called with a HeteroExplanation they failed, masks live per node/edge type and complement masks need to be computed per-type.

Changes

• fidelity.py: early dispatch to a new _hetero_fidelity helper for HeteroExplanation; the homogeneous body is left unchanged. Adds a private _hetero_model_kwargs helper that reconstructs the original model kwargs from _model_args, respecting the *_dict suffix convention used by the explainer.

• faithfulness.py: same dispatch pattern with _hetero_unfaithfulness. In the heterogeneous branch top_k is applied per node type and clamped to the number of features available for that type.

• New tests test_fidelity_hetero (2) and test_unfaithfulness_hetero (24, parametrised on top_k × explanation_type × node_mask_type × return_type).

Dependency

Full kwargs propagation (e.g. edge_attr_dict) requires _model_args to be populated on HeteroExplanation, which is the subject of #10676 . This PR uses getattr(explanation, '_model_args', []) so it degrades gracefully when running against a base without that change.