Switch label_misinformation to Mistral by default

[gmguarino]

Summary

• Adds MistralSinglePromptPipeline to jobs/label_misinformation alongside the existing OpenAI SinglePromptPipeline, selectable via a new LLM_PROVIDER env var (default: mistral). Uses mistral-small-2603 with a new French structured prompt (reasoning + misinformation bool, prompt version "0.1.0"), hardened against a repetition-loop failure mode found during evaluation (penalties, stop sequence, retry-on-parse-failure).
• Whether MIN_MISINFORMATION_SCORE thresholding applies is now a property of the active prompt (DisinformationPrompt.binary), not the provider — binary prompts (the new Mistral one) filter on score > 0 directly; continuous prompts (the existing OpenAI one) keep using the threshold.
• Adds a standalone OpenAI eval script (climateguard/finetuning/api/openai/completions/test.py) mirroring the existing Mistral one, used to compare the two providers on the DataForGood/climateguard train/test splits before choosing Mistral.
• Fixes an "Event loop is closed" bug in the Mistral eval script (client was reused across separate asyncio.run() calls).
• Documents all of the above in jobs/README.md.

Compatibility / rollout checklist

For this repo/branch to work as-is (Mistral, the new default):

• Create jobs/label_misinformation/secrets/mistral_key locally (and the equivalent secret in whatever deploys this job — Vaultwarden/CI/Scaleway) with a real Mistral API key. Not included in this PR since it's a secret.
• Confirm MISTRAL_API_KEY is available wherever docker-compose.yml's app/test/testconsole services run (the mistral_key docker secret is already wired up in this PR).
• Run docker compose up testconsole -d && docker compose exec testconsole pytest tests/test_pipeline.py tests/test_prompts.py -vv to sanity-check the new pure-function tests pass.
• Smoke-test docker compose up app against a small known day/channel and confirm in the logs it used the Mistral provider, and that flagged rows have sensible French model_reason text.

For any job/service that should keep using OpenAI models (nothing removed — OpenAI support is fully intact as an opt-in path):

• Set LLM_PROVIDER=openai explicitly in that service's environment (main.py defaults to mistral otherwise).
• Ensure OPENAI_API_KEY / the openai_key docker secret is still populated (unchanged from before this PR).
• No prompt config needed — the OpenAI path always uses prompt version "0.0.1" (unchanged, English, plain 0-10 score) regardless of PIPELINE_PRODUCTION_PROMPT.
• MIN_MISINFORMATION_SCORE continues to apply exactly as before for this path (untouched behavior).
• MODEL_NAME/MODEL_NAME_BRAZIL should be set to an OpenAI model name (e.g. gpt-4o-mini) for that service, since country.py's defaults were changed to mistral-small-2603 to match the new default provider.

Out of scope (left untouched, per explicit request): the shadow docker-compose service and requirements_shadow.txt (BERT shadow pipeline) — still OpenAI-only, no LLM_PROVIDER wiring added there.

Test plan

• parse_structured_response unit tests (valid JSON, malformed-JSON-with-regex-fallback, unparseable-defaults-to-zero) run locally and pass.
• Existing parse_response/parse_response_reason tests re-verified locally, still pass.
• test_prompts.py's existing assertions re-verified against the new default (PIPELINE_PRODUCTION_PROMPT now resolves to "0.1.0", prod=True).
• All touched files pass py_compile.
• Full docker compose up test / app smoke test (needs live secrets/DB — not run in this environment).