fix(ai-red-teaming): collect multimodal attack results via assessment.run()

[rdheekonda]

Summary

The generated multimodal workflow reported 0 finished trials (and wrote no analytics) even when the attack ran successfully and traced to the platform.

Root cause: the multimodal template ran each attack with await attack.run() directly on the Study. Only Assessment.run(study) appends the AttackResult to assessment.attack_results (assessment.py). Running the Study directly left that list empty, so _write_local_analytics saw no results → "0 finished trials". Every non-multimodal generator already uses assessment.run(...).

Discovered while probing a real SageMaker Gemma3-vision endpoint: trials executed, judge scored them, yet the CLI showed 0 trials.

Changes

• await attack.run() → await assessment.run(attack) in the multimodal main template (collects + uploads the result; also feeds local analytics).
• Fix per-set except handler to use the real loop var set_number instead of undefined i (which raised NameError inside the handler, masking real errors).
• Regression tests (proven red→green): assert the generated script uses assessment.run(attack) and not attack.run(), and that the handler uses set_number.
• Fix stale test_per_media_prompts_list assertion (PROMPTS[i] → set_prompt = PROMPTS[media_idx] after a prior refactor).
• Bump capability 1.8.0 → 1.9.0 (minor).

Testing

• TestGenerateMultimodalAttack: 17 passed (editable SDK 2.0.35).
• Regression tests fail on the pre-fix template, pass after.
• End-to-end re-run vs the live SageMaker endpoint: analytics now written; 2 trials; ASR reported.

Note

data/mmsafety_illegal_0.png in the workspace is a JPEG mislabeled .png; this is a data-hygiene issue, not the cause here (the workflow re-encodes media via PIL). The 0-trials symptom was purely the attack.run() bug.