feat: add MiniMax provider support for training data generation

[octo-patch]

Summary

Add MiniMax as a third LLM backend option for generating training data in gendata.py, alongside Ollama and Claude API.

Changes

• New --minimax flag for gendata.py to use MiniMax API (OpenAI-compatible endpoint)
• Supported models: MiniMax-M2.7 (default) and MiniMax-M2.7-highspeed
• No new dependencies — uses urllib.request (same pattern as Ollama integration)
• Chain-of-thought handling — strips <think> tags from MiniMax responses
• Temperature clamping — enforces MiniMax's valid range (0.0, 1.0]
• Environment variables: MINIMAX_API_KEY (required), MINIMAX_BASE_URL (optional)
• Unit tests: 20 tests including real API integration test
• Documentation: Updated README.md and examples/guess/README.md

Usage

# Set API key
export MINIMAX_API_KEY="your-key"

# Generate training data with MiniMax
./gendata.py --topic elephant --minimax -n 500 > training.txt

# Use highspeed model
./gendata.py --topic elephant --minimax --model MiniMax-M2.7-highspeed -n 500

API Reference

• Chat (OpenAI Compatible): https://platform.minimax.io/docs/api-reference/text-openai-api

Test Plan

• All 20 unit tests pass (provider config, URL construction, auth headers, temperature clamping, think-tag stripping, error handling)
• Integration test passes with real MiniMax API
• End-to-end CLI test generates valid training data (--minimax flag)
• Existing Ollama and Claude backends remain unaffected
• python3 -c "import ast; ast.parse(open('examples/guess/gendata.py').read())" passes