test: parametrise canary test across attacks (qmia, lira, worstcase)

[ssrhaso]

Is your feature request related to a problem? Please describe.
test_qmia_predicts_canaries in tests/attacks/test_qmia_attack.py (added in #435) is qmia-specific. lira and worstcase have no equivalent positive-signal test that verifies the attack actually flags deliberately-vulnerable records. duplicating the canary scaffolding (data generation, 9-nn boundary selection, label flip, target build) per attack would be churn, and per-attack thresholds would drift out of sync over time.

Describe the solution you'd like
follow-up from #435: @jim-smith suggested lifting test_qmia_predicts_canaries into a generic test_attack_predicts_canaries(attack_cls, tmp_path) parametrised across [QMIAAttack, LIRAAttack, WorstCaseAttack].

Describe alternatives you've considered
leave it qmia-specific and write separate canary tests for each attack ad hoc. rejected: duplicates the canary scaffolding, encourages thresholds to drift, and means the next attack added to the project starts without a positive-signal test.

Additional context
notes from the qmia version that may matter for lira/worstcase:

• bootstrap=False was needed for the rf to actually memorise the canaries; default rf (bootstrap=True) dilutes the signal across trees that never saw the canary.
• ranking canaries within the training slice was bimodal across seeds because the qmia quantile threshold extrapolates at rare label/location combos. the stable metric was canary-vs-test auc.
• review discussion: feat: add QMIAAttack quantile regression membership inference attack #435

would consider extracting the canary fixture into a shared helper, then pytest.parametrize over the three attacks with per-attack auc thresholds - since shadow-model attacks will likely need looser bounds than qmia.

[ssrhaso]

@jim-smith please update the issue as you see fit, this is my understanding of the issue

[jim-smith]

@jim-smith please update the issue as you see fit, this is my understanding of the issue

@ssrhaso I think you described it well. thanks

[jim-smith]

closed by #462