feat(airt): capture generated media output in multimodal workflows

[rdheekonda]

Summary

Follow-up to the SDK multimodal-output work (dreadnode-tiger #1891). Enables the
platform-generated multimodal workflow to capture what a target generates,
not just what it says.

• The generated @task target now returns the model's full reply when it
  contains generated media (image / audio / video), so those output parts are
  captured and rendered in findings & traces. Plain text replies still return
  .content (a str), so text-out targets are unchanged.
• This is the piece that lets media-out targets flow through the capability UI:
  an image-out model, or a speech-to-speech target such as Amazon Nova Sonic
  driven via a custom target (audio-in → audio-out).
• Bumps capability 1.5.0 → 1.6.0.

Why this is safe

The change is gated on content_parts containing an image/audio/video kind.
Text-only replies take the identical reply.content path as before — no change
to the overwhelming majority (text-out) flow. Added a regression test asserting
both the media-kind guard and the text fallback.

Sequencing

Merge alongside dreadnode-tiger #1891 (SDK: media-out response_to_parts +
per-modality scoring). The SDK's response_to_parts is what turns the returned
reply into typed output parts. Per-modality scoring of generated media is a
separate future step (the current llm_judge stringifies its input, so it can't
yet judge a generated image/audio — that needs a multimodal judge scorer).

Test plan

• pytest capabilities/ai-red-teaming/tests/test_attack_runner.py — all pass
• pytest tests/test_capability_release_plan.py — passes with 1.6.0
• Post-merge: run a media-out target through the capability, confirm output
  image/audio/video renders & plays in Overview + Traces