feat: add MiniMax as alternative LLM provider for data preprocessing

[octo-patch]

Summary

Add configurable LLM provider support to the data preprocessing captioning scripts. Users can now use MiniMax (MiniMax-M2.7, MiniMax-M2.5) as an alternative to OpenAI GPT-4V for frame and video captioning, or point to any OpenAI-compatible API.

Changes

• New file data_preprocess/llm_provider.py: Shared provider configuration module with PROVIDER_PRESETS, create_client(), get_model_name(), and clamp_temperature()
• Modified 3 captioning scripts: step2_1_GPT4V_frame_caption.py, step3_1_GPT4V_video_caption_concise.py, step3_1_GPT4V_video_caption_detail.py — all now accept --provider, --base_url, --model arguments
• Updated run.sh: Added PROVIDER variable for easy switching
• Updated README: Added LLM Provider section with MiniMax usage docs
• 9 files changed, 556 additions, 36 deletions

Usage

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

# Use MiniMax for frame captioning
python data_preprocess/step2_1_GPT4V_frame_caption.py \
    --provider minimax \
    --image_directories ./step_1

# Use MiniMax for video captioning
python data_preprocess/step3_1_GPT4V_video_caption_concise.py \
    --provider minimax \
    --input_file ./2_2_final_useful_gpt_frames_caption.json

Backward Compatibility

• Default provider remains openai — no changes needed for existing workflows
• The --api_key argument still works as before
• Environment variable OPENAI_API_KEY is used as fallback for any provider

Test Plan

• 26 unit tests covering provider presets, argument parsing, client creation, model resolution, temperature clamping, and function signatures
• 3 integration tests verifying actual MiniMax API connectivity for text captioning
• All 29 tests passing