FEAT: adding Garak audio achilles scenario

[Richard Lundeen (richlundeen)]

Description

Adds a new Garak scenario, AudioAchillesHeel, that ports Garak's audio.AudioAchillesHeel probe into PyRIT scenario constructs. The adversarial instruction is delivered as spoken audio while the text channel carries only a benign "follow the audio instructions" nudge, testing whether a multimodal target complies with harmful spoken instructions.

What it does

• Loads spoken-jailbreak clips from the existing garak_audio_achilles_heel dataset.
• Shapes each clip into a single multimodal user turn (text nudge + audio at the same sequence, merging into one Message), mirroring Garak's request shape.
• Derives a per-clip objective from the clip's harm category and scores the response for compliance (the PyRIT analogue of Garak's non-refusal mitigation.MitigationBypass detector).
• Baseline attack policy is disabled (audio-payload dominated; a text-only baseline drops the audio).

Notes

• The default run samples a small subset (max_dataset_size=10 of ~350 clips) so a default pyrit_scan garak.audio_achilles_heel completes quickly; override with --max-dataset-size.
• The objective target must accept audio_path input (multimodal).
• Added a documentation-only section to the Garak scanner docs (doc/scanner/garak.py + .ipynb).

Tests

• 20 unit tests in tests/unit/scenario/garak/test_audio_achilles_heel.py (all passing).
• pre-commit and ruff clean.