Add MiniMax as LLM evaluation provider

[octo-patch]

Summary

• Add MiniMax (M2.5, M2.5-highspeed, M2.7) as a new LLM provider for the LoCoMo long-term conversational memory benchmark
• Follow the existing provider pattern (GPT, Claude, Gemini) with run_minimax() in global_methods.py and task_eval/minimax_utils.py
• MiniMax models offer 204K-1M context windows via OpenAI-compatible API, making them suitable for long-term conversation evaluation

Changes

File	Description
global_methods.py	Add run_minimax() and set_minimax_key() with OpenAI-compatible API, temperature clamping, think-tag stripping
task_eval/minimax_utils.py	New MiniMax evaluation utils with get_minimax_answers() following claude_utils.py pattern
task_eval/evaluate_qa.py	Add MiniMax model dispatch alongside GPT/Claude/Gemini
scripts/evaluate_minimax.sh	Evaluation script for MiniMax-M2.5 and MiniMax-M2.5-highspeed
scripts/env.sh	Add MINIMAX_API_KEY environment variable
README.MD	Add MiniMax evaluation instructions
tests/test_minimax_unit.py	27 unit tests covering API calls, model mapping, temperature clamping, think-tag stripping, answer parsing
tests/test_minimax_integration.py	3 integration tests with real MiniMax API calls

Usage

Set your MiniMax API key in scripts/env.sh, then run:
bash scripts/evaluate_minimax.sh

Test plan

• 27 unit tests passing (mocked API calls)
• 3 integration tests passing (real API calls)
• Verify evaluation output format matches existing providers