Add Inspect AI evaluation package

[emmepra]

Summary

Adds an installable Inspect AI package for AHB, including dataset mapping, the ahb @task entrypoint, the public safe/unsafe scorer, tests, README usage for OpenAI/OpenRouter providers, CI test coverage, and ignored Inspect log outputs.

Why

AHB needs an upstream Inspect-ready package before it can be submitted through the Inspect Evals Register. This keeps icaro-lab/ahb as the canonical public repo and avoids a separate ahb-inspect repository.

Linked Issue

Closes #1

Verification

• Local: uv sync --python 3.11 --extra test; uv run python -m compileall ahb_eval scripts src tests; uv run pytest; uv run inspect eval ahb_inspect/ahb --model mockllm/model --model-role grader=mockllm/model --limit 1 --no-log-realtime; uv run inspect eval ahb_inspect/ahb --model openrouter/openai/gpt-oss-20b --model-role grader=openrouter/openai/gpt-oss-20b --limit 1 --max-tokens 128 -T max_judge_tokens=128 --no-log-realtime (attack_success_rate=0.000)
• Browser: Not applicable; no frontend behavior changed.
• Browser evidence: Not applicable.
• CI: Previously passed on PR Add Inspect AI evaluation package #2; pending refresh after follow-up provider dependency commit.

Auto-Merge Eligibility

Not requested. This is a normal PR to dev; local checks and one real OpenRouter smoke pass, and the PR is left open for user review before merge.

Blueprint Updates

No versioned blueprint update. The repo has no .project/blueprint and .project remains local-only for this public repo; the change follows AGENTS.md public release constraints.

Risks / Rollback

The scorer depends on current inspect-ai APIs and is pinned at inspect-ai>=0.3.220. OpenAI-compatible providers require openai>=1.50.0, now included. Rollback is a normal revert of this package layer; no dataset publication or private methodology is changed.

Follow-ups

After merge, replace the placeholder SHA in the local Register YAML draft and open a separate PR to UKGovernmentBEIS/inspect_evals under register/ahb/eval.yaml. Larger paid model validation can be run only after explicit approval.