Add Mistral provider support

[meaningfool]

Summary

• add direct mistral service support using Mistral's OpenAI-compatible API
• document the new service alias and required env vars in the README
• fix --judge-model so the selected Claude judge model is actually used
• make the comprehensive eval helper usable for unstable providers by supporting sequential retries and skipping incomplete runs
• fix the run-directory race by returning the created run dir directly from the CLI run path

Mistral Config Used For Benchmarking

• model: mistral-small-2603
• endpoint: https://api.mistral.ai/v1
• interface: OpenAI-compatible /v1/chat/completions
• benchmark used provider defaults for sampling and tool behavior: no explicit temperature, top_p, seed, tool_choice, or parallel_tool_calls overrides
• later stable runs used MTE_TEXT_IDLE_TIMEOUT_SECS=240 in the harness to avoid killing slow turns; this is a runner setting, not a model setting

Validation

• uv run python -m py_compile scripts/run_comprehensive_eval.py src/multi_turn_eval/cli.py src/multi_turn_eval/judging/claude_judge.py src/multi_turn_eval/pipelines/base.py
• smoke and full Mistral benchmark runs completed successfully with Opus judging