feat: add MiniMax Cloud API as alternative prompt enhancement provider

[octo-patch]

Summary

Add MiniMax M2.7 as a cloud-based prompt enhancement option for users who don't have enough GPU memory to run the local reprompt model (32B or smaller).

This offloads prompt rewriting to MiniMax's OpenAI-compatible API while keeping the exact same predict()/to() interface as the local RePrompt models, so users only need to change one parameter:

pipe = HunyuanImagePipeline.from_pretrained(
    model_name="hunyuanimage-v2.1",
    reprompt_model="minimax",  # cloud prompt enhancement
    use_fp8=True,
)
image = pipe(prompt="A cute penguin", use_reprompt=True)

Changes

• New file: hyimage/models/reprompt/reprompt_cloud.py — RePromptCloud class
  • OpenAI-compatible SDK calls to MiniMax API
  • Think-tag stripping + <answer> tag parsing (matches local model output format)
  • Temperature clamped to (0, 1] per MiniMax API constraint
  • Graceful fallback to original prompt on API errors
• Modified: hyimage/models/model_zoo.py — HUNYUANIMAGE_REPROMPT_MINIMAX() factory
• Modified: hyimage/diffusion/pipelines/hunyuanimage_pipeline.py
  • _resolve_reprompt_config() for clean model name → config mapping
  • Cloud-aware _load_reprompt_model() (skips models_root_path for cloud models)
  • Updated from_pretrained() docstring
• Modified: README.md — Cloud-based prompt enhancement usage guide
• New file: tests/test_reprompt_cloud.py — 36 unit + 3 integration tests

5 files changed, 717 additions, 4 deletions

Environment Variables

Variable	Required	Default
MINIMAX_API_KEY	Yes	—
MINIMAX_MODEL	No	MiniMax-M2.7
MINIMAX_BASE_URL	No	https://api.minimax.io/v1

Test Plan

• 36 unit tests pass (mocked API, no GPU/network needed)
• 3 integration tests pass (live MiniMax API calls)
• Tests cover: init, env vars, predict, think-tag stripping, answer-tag parsing, error fallback, Chinese prompts, interface compatibility
• Existing pipeline behavior unchanged (no modifications to local reprompt flow)