Add MiniMax as direct LLM provider (M2.7 + M2.7-highspeed)

[octo-patch]

Summary

• Adds MiniMax as a first-class provider alongside OpenRouter and OpenAI, enabling direct API access to MiniMax models without routing through OpenRouter
• Introduces MiniMaxClient with temperature clamping to (0.0, 1.0], <think> tag stripping for M2.5+ models, and minimax/ prefix stripping
• Upgrades MiniMax from M2.5 to M2.7 + M2.7-highspeed (204K context) in both config.json and config.v2.json
• Adds wildcard provider routing (minimax/* to minimax) in config model_providers
• Wires up MINIMAX_API_KEY env var in collect, grade, and grade_panel flows
• Updates README with MiniMax setup instructions

Changes

File	Change
scripts/openrouter_benchmark.py	MiniMaxClient, MiniMaxAPIError, helper functions, provider registration, client wiring
config.json	M2.7 + M2.7-highspeed models, minimax/* provider routing
config.v2.json	Same as above for v2 config
README.md	MINIMAX_API_KEY env var, minimax routing example
tests/test_minimax_provider.py	30 unit tests + 3 config tests + 3 integration tests

Test plan

• 47 tests pass (python3 -m pytest tests/test_minimax_provider.py -v)
• Run dry-run end-to-end to verify config parsing
• Set MINIMAX_API_KEY and run a single-model collect to verify end-to-end