Improve coverage for embedding evaluation helpers and fallback branches

[Copilot]

This PR addresses the coverage gap in embedding evaluation utilities by adding focused unit coverage for untested operator paths, empty-input handling, and error/fallback behavior. The goal is to harden py3plex/ml/embedding/evaluation.py against edge-case regressions.

• Coverage target

  • Added a dedicated test module: tests/test_ml_embedding_evaluation.py.
  • Focuses specifically on branch-heavy helpers in py3plex/ml/embedding/evaluation.py.

• Operator and feature-construction branches

  • Added explicit tests for edge_operator modes: hadamard, average, weighted_l1, weighted_l2, cosine.
  • Added invalid-operator assertion path.
  • Added empty-edge behavior checks in build_edge_features for both vector and scalar-output operators.

• Evaluation helper paths

  • Added direct helper coverage for:
    • evaluate_link_prediction
    • evaluate_node_classification
    • evaluate_clustering
  • Extended evaluate_link_prediction_embeddings coverage for both valid classification input and single-class failure behavior.

• Fallback/error propagation in split logic

  • Added targeted coverage for evaluate_node_classification_report:
    • first split attempt fails, fallback succeeds
    • both attempts fail and propagate the error

split_mock = Mock(
    side_effect=[
        ValueError("forced split failure"),
        (X_train, X_test, y_train, y_test),
    ]
)
monkeypatch.setattr(evaluation, "train_test_split", split_mock)

report = evaluation.evaluate_node_classification_report(
    np.zeros((6, 2), dtype=np.float32),
    [0, 1, 0, 1, 0, 1],
    random_state=0,
)