Fix non sensitivity

[Tprojects66554]

Description

This pull request addresses two critical issues in the Non-Sensitivity metric related to the features_in_step parameter and the internal logic of pixel perturbations.

1. Logical inconsistency in perturbation evaluation
   When features_in_step > 1, multiple pixels (both feature and non-feature) are perturbed simultaneously within the same step.
   As a result, the computed difference between the perturbed and original predictions (y_pred) corresponds to mixed pixel groups, making it impossible to determine whether individual pixels preserve the model’s insensitivity property.

2. Shape mismatch causing ValueError
   When running the metric with features_in_step != 1, Quantus raised:

   ValueError: operands could not be broadcast together with shapes (17,24) (17,150528)
   

   This occurred at:

   return (preds_differences ^ non_features).sum(-1)

   where:

   • non_features → shape (batch_size, n_features)
   • preds_differences → shape (batch_size, n_perturbations)
     Since n_perturbations != n_features for multi-step perturbations, the XOR (^) operation failed due to incompatible dimensions.

These issues caused both logical misinterpretation of sensitivity violations and runtime failures during evaluation.
Link to the issue

• https://github.com/understandable-machine-intelligence-lab/Quantus/issues/367

Implemented changes

• Rewrote the perturbation loop to evaluate per-pixel sensitivity, ensuring clean separation between feature and non-feature perturbations.
• Adjusted the accumulation logic to correctly track prediction stability across perturbation steps.
• Fixed the broadcasting mismatch by aligning array dimensions and reshaping operations before computing sensitivity violations.
• Added debug information and improved documentation for reproducibility and clarity.

Minimum acceptance criteria

• All tests under tests/metrics/test_non_sensitivity_metric.py and related evaluation modules pass successfully across supported environments (py310–py311).
• The metric produces consistent scores for all features_in_step configurations without shape or logic errors.
• Reviewer confirmation by @annahedstroem or a Quantus core maintainer.

[Tprojects66554]

Hi @annahedstroem I would appreciate it if you could take a look at the PR to see what you think about it.