Allow Infidelity to accept all-negative attributions

[delaidam]

Infidelity computes a_batch * (x - x_perturbed) summed over features (see evaluate_batch), a signed dot product with no mathematical requirement for non-negative attributions (see the referenced paper's formula, https://arxiv.org/pdf/1901.09392). The shared assert_attributions() check used by all metrics via Metric.general_preprocess()/explain_batch() rejected all-negative attributions unconditionally, which is overly constraining for Infidelity specifically (e.g. SHAP explanations for a binary classifier, where the sign indicates which class is promoted).

Adds Metric.allow_negative_attributions (default False, preserving current behavior for every other metric) and a check_all_negative parameter on assert_attributions(), set by Infidelity to skip just that one check.

Verified directly (pytest could not run end-to-end in this environment -- tests/conftest.py's autouse cifar10.load_data() fixture requires network access that isn't available here):

• assert_attributions() rejects all-negative by default, allows it with check_all_negative=False
• Infidelity.allow_negative_attributions is True, other metrics (checked: FaithfulnessCorrelation) remain False
• Infidelity(...)(...) runs end-to-end on a tiny synthetic model with an all-negative precomputed a_batch and returns valid scores, where it previously raised AssertionError before evaluation started

Closes #375

Description

• Insert a clear and concise description of what the PR is about.
• Optionally, make a reference to an existing issue.
• Describe the expected improvements of Quantus that results from this PR. Why was this work completed?

Implemented changes

• Insert a description of the changes implemented in the pull request.
  • Change 1
  • Change 2
  • Change 3

Minimum acceptance criteria

• Specify what is necessary for the PR to be merged with the main branch.
• @mentions of the person that is apt to review these changes e.g., @annahedstroem