Add Inspect backends for MTBench

[adamhrin-amd]

Summary

• Add endpoint-backed Inspect tasks for English mt_bench_inspect and Finnish mt_bench_fi_inspect.
• Register English alongside the existing native mt_bench task, while keeping native mt_bench on the existing flowaicom/Flow-Judge-v0.1 vLLM judge path.
• Register Finnish as a built-in task via mt_bench_fi/main.py, backed by LumiOpen/mtbench_multi, subset fi, split test.
• Run both Inspect tasks as true two-turn MTBench tasks: generate turn 1, append turn 2, generate turn 2, then score turn_1 and turn_2 separately.
• Use an OpenAI-compatible scorer endpoint for judging via SCORER_MODEL_BASE_URL, SCORER_MODEL_PATH, and VLLM_API_KEY.
• Preserve MTBench reference answers when present. Finnish scoring reuses the existing Flow-Judge MTBench prompt templates and adds only a small Finnish-content note.
• Validate scorer endpoint configuration in the Inspect solver before generation, while creating the scorer model lazily so registry loading and native task imports do not require scorer env vars.

Verification

• ruff check src/lighteval/tasks/tasks/mt_bench/main_inspect.py src/lighteval/tasks/tasks/mt_bench_fi/main.py src/lighteval/tasks/tasks/mt_bench_fi/main_inspect.py src/lighteval/main_inspect.py
• python3 -m py_compile src/lighteval/tasks/tasks/mt_bench/main_inspect.py src/lighteval/tasks/tasks/mt_bench_fi/main.py src/lighteval/tasks/tasks/mt_bench_fi/main_inspect.py src/lighteval/main_inspect.py
• TensorWave Qwen3.5-4B smokes completed successfully with /shared_silo/scratch/models/Qwen3.5-4B as target and judge:
  • English mt_bench_inspect: job 38686, dashboard eval 927.
  • Finnish mt_bench_fi_inspect: job 38687, dashboard eval 928.
• The smokes used runtime --reasoning-tags set to [('<think>', '</think>'), ('', '</think>')]; raw logs observed thinking traces before stripping, and scratch assertions confirmed scorer answers and judge prompts were tag-free for both turns.

Notes

• The successful TensorWave smokes used the current TensorBase MTBench wrapper shape: no --custom-tasks, compatible vLLM020 packages, and scorer endpoint env vars exported by the wrapper.
• The Finnish dataset loaded as 80 test rows from LumiOpen/mtbench_multi / fi during the smoke.

[ezosa]

Approved. But we might want to add a language checking step to MTBench-FI so that only responses classified as Finnish by a language identifier (e.g. langdetect) will be judged by the LLM. From previous experience, LLMs are not always consistent when asked to identify the language of the response.

[adamhrin-amd]

Okay! Merging this one and will open a new work for the lang detection