fix: report generalisation gap instead of inverted accuracy gap

[shamykyzer]

As reported in #468, StructuralAttack's generalisation metric was inverted and mislabelled: it returned the train/test accuracy gap with the sign flipped, so an overfit model showed a negative value.

This changes it to report the generalisation gap = test error - train error (for classification, (1 - test_acc) - (1 - train_acc)), which is positive when the model overfits. Classification only; regression is left to #414 as discussed on the issue.

• Flip the sign in the SklearnModel and PytorchModel generalisation methods.
• Clarify the docstrings, and fix the swapped train_acc/test_acc descriptions in StructuralAttackResults.
• Update the affected test assertions; add a CHANGELOG entry.

Closes #468

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 99.66%. Comparing base (b7885f1) to head (9c05bfe).

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #472   +/-   ##
=======================================
  Coverage   99.66%   99.66%           
=======================================
  Files          28       28           
  Lines        3633     3633           
=======================================
  Hits         3621     3621           
  Misses         12       12           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[rpreen]

Thanks for this. Please can you remove any references to issue numbers to keep the source clean? Also the variable/function names (pretty much anywhere that deals with "generalisation error") like generalisation_error need to be renamed to generalisation gap for clarity.