Add Harbor adapter for judge v2

[xeophon]

Adds the Harbor adapter changes to the judge-v2 branch.

References/dependencies:

• Ports Harbor support from Add Harbor Framework Support  #8.
• Keeps the Prime VM/GPU verifier assumptions aligned with APR-262 Add Harbor verifiers support for v1 PrimeIntellect-ai/verifiers#1401.

What changed:

• Adds src/harbor_adapter for generating 28 Harbor tasks across the 7 PostTrainBench benchmarks and 4 base models.
• Uses Prime driver-aware CUDA images for the agent and separate verifier containers.
• Ports the verifier script to PR New judging system (v2) #50's judge-v2 flow: canonical judgement_gpt5_4.json, archival judgement_api.json, trace parsing, and reward zeroing on contamination/base-model violations.
• Defaults verifier OpenAI-compatible calls to Pinference through OPENAI_BASE_URL and CODEX_MODEL; forwards OPENAI_BASE_URL to agent env only for ArenaHard/HealthBench evaluator-judge runs.
• Updates README wording now that Harbor task generation exists.

I could not push directly to aisa-group:new_judge_v2: GitHub rejected the xeophon token with 403, and PR #50 has maintainerCanModify=false. This PR targets that branch so it can be merged into PR #50.

Tested:

• bash -n src/harbor_adapter/template/environment/entrypoint.sh src/harbor_adapter/template/environment/system_monitor.sh src/harbor_adapter/template/tests/test.sh
• uv run --no-project python3 -m py_compile src/harbor_adapter/adapter.py src/harbor_adapter/run_adapter.py
• git diff --check
• uv run --no-project python3 src/harbor_adapter/run_adapter.py --benchmark gsm8k --model qwen3-1.7b --output /tmp/ptb-harbor-smoke
• uv run --no-project python3 src/harbor_adapter/run_adapter.py --benchmark healthbench --model qwen3-1.7b --output /tmp/ptb-harbor-smoke
• uv run --no-project python3 src/harbor_adapter/run_adapter.py --list

[xeophon]

ahhhh sorry