refactor(discriminatory-behavior): migrate to new defense architecture

[asim29]

Context

Sixth in the refactor/risk-modules stack. Stacked on #78 (unauth model ownership vertical).

Changes

amulet/discriminatory_behavior/defenses/adversarial_debiasing.py

• AdversarialDebiasing subclasses DiscriminatoryBehaviorDefense(ABC) and delegates model, criterion, optimizer, train_loader, test_loader, device to super().__init__().
• Dead imports (numpy, sklearn.metrics) removed — no longer used after adversary_auc was moved to the metrics class.

amulet/discriminatory_behavior/metrics/discriminatory_behavior.py

• adversary_auc() extracted from the defense into DiscriminatoryBehavior as a @staticmethod, keeping attacks/defenses free of metric logic per API contract.
• p_rule() now returns 0.0 on divide-by-zero instead of propagating nan.
• Docstrings converted to Google style; RST markup removed.

examples/defense_pipelines/run_adversarial_debiasing.py

• Updated to use load_or_train instead of manual save/load blocks.

tests/integration/test_discriminatory_behavior.py (new)

• Smoke test for AdversarialDebiasing.train_fair() verifying model weights change and discmodel attribute is set.

tests/unit/test_metrics_discriminatory.py (new)

• Unit tests for p_rule, demographic_parity, true_positive_parity, false_positive_parity, and accuracy covering edge cases including divide-by-zero.

Test plan

• uv run pytest tests/integration/test_discriminatory_behavior.py -v -m integration
• uv run pytest tests/unit/test_metrics_discriminatory.py -v
• uv run pre-commit run --all-files
• CI passes on this branch

[asim29]

Fixes #59 and #60