feat: reduce JSON report size (#437)

[ssrhaso]

Summary

Reduce JSON report file size by removing large derived arrays and externalising individual record data.

• Strip fpr, tpr, and roc_thresh arrays from JSON serialisation via a _json_exclude_keys class attribute on the Attack base class. PDF generation is unaffected as it reads from the in-memory dict before serialisation.
• Externalise LiRA individual record scores to a compressed .npz file, storing a relative filename in JSON for portability. Includes y_pred_proba and y_test to allow ROC recomputation from stored data.
• Add backwards compatibility guard in LogLogROCModule for JSON files without ROC arrays.

For a CIFAR10-scale dataset, this reduces a single attack JSON from approximately 1.2MB to under 100KB.

Closes #437

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 99.52%. Comparing base (1863c65) to head (2b70e0e).

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #440   +/-   ##
=======================================
  Coverage   99.51%   99.52%           
=======================================
  Files          23       23           
  Lines        2692     2713   +21     
=======================================
+ Hits         2679     2700   +21     
  Misses         13       13           

☔ View full report in Codecov by Sentry.
📢 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.

[ssrhaso]

@rpreen Please let me know what you think of this when you get a chance!

[jim-smith]

@ssrhaso does this affect the meta-attack?

[ssrhaso]

@ssrhaso does this affect the meta-attack?

@jim-smith To my understanding and after checking thoroughly, no it does not.

MetaAttack reads per-record scores from the in-memory sub-attack object, not from the JSON, so the key-stripping doesn't affect it.

Spotted one related side-effect while checking: the new .npz write in LIRAAttack._save_attack_metrics was gated on report_individual only, so MetaAttack sub-runs (which set write_report=False) would have dropped stray lira_individual.npz files in each sub_dir. Pushed a fix to also gate on write_report.