feat: add MiniMax as generation backend

[octo-patch]

Summary

Add MiniMax as a first-class LLM generation backend alongside the existing vllm, openai, and hf backends.

MiniMax provides OpenAI-compatible cloud APIs with models featuring up to 1M context windows, making them well-suited for RAG workloads that require processing large retrieved contexts.

Changes

• servers/generation/src/generation.py — Added minimax backend with:
  • Auto-detection of MINIMAX_API_KEY environment variable
  • Temperature clamping to MiniMax's accepted (0, 1] range
  • Automatic <think>...</think> tag stripping (configurable via strip_think_tags)
  • Default model: MiniMax-M2.7 (1M context)
  • Concurrent request support with exponential backoff retry
  • Two static helper methods: _clamp_temperature() and _strip_think_tags()
• servers/generation/parameter.yaml — Added MiniMax config section with all available options
• examples/minimax_rag.yaml — Example RAG pipeline using MiniMax backend
• examples/parameter/minimax_generation_parameter.yaml — Full parameter reference
• README.md / docs/README_zh.md — Added "Supported Cloud LLM Backends" table documenting all four backends with MiniMax usage instructions
• tests/test_minimax_generation.py — 38 unit tests covering temperature clamping, think-tag stripping, initialization, and generation
• tests/test_minimax_integration.py — 3 integration tests (auto-skipped when MINIMAX_API_KEY is not set)

Supported Models

Model	Context	Notes
MiniMax-M2.7	1M tokens	Latest, default
MiniMax-M2.7-highspeed	1M tokens	Fast variant
MiniMax-M2.5	256K tokens	Previous generation
MiniMax-M2.5-highspeed	204K tokens	Fast, long context

Usage

export MINIMAX_API_KEY="your-api-key"
ultrarag run examples/minimax_rag.yaml

Or set backend: minimax in your generation parameter file.

Test Plan

• 38 unit tests pass (temperature clamping, think-tag stripping, init validation, mock generation)
• 3 integration tests pass against live MiniMax API
• Verify no regression on existing vllm/openai/hf backends

9 files changed, 857 additions(+), 3 deletions(-)

[xhd0728]

@octo-patch Thanks for the PR~

The main issue is that this introduces provider-specific logic and a dedicated parameter setup for MiniMax in the generation framework. We'd prefer not to maintain a separate parameter path for one specific provider, and instead keep the backend interface and configuration as unified as possible.

Therefore, we won't merge this PR in its current form.