feat: add MiniMax as LLM provider in caption annotation pipeline

[octo-patch]

Summary

This PR adds MiniMax as a supported LLM provider in the SpatialVID caption annotation pipeline.

Changes

• caption/utils/api_call.py: Detect api.minimax.io in base_domain and route to MiniMax's OpenAI-compatible endpoint (/v1/chat/completions). Temperature is kept at 0.1, within MiniMax's required (0.0, 1.0] range. The duplicated response-extraction branches are unified into a single return path.
• caption/README.md: Add a Supported LLM Providers table listing MiniMax, Qwen, and Gemini with their --base_domain values, plus concrete MiniMax usage examples for LLM/inference.py and tagging/inference.py.
• caption/tests/test_api_call.py: 12 unit tests (mock-based) + 3 integration tests against the live MiniMax API, covering endpoint routing, temperature constraints, auth headers, multimodal list content, and backward compatibility with Qwen/Gemini.

Usage

# LLM Captioning with MiniMax M2.7
python caption/LLM/inference.py \
  --csv_path path/to/data.csv \
  --pose_load_dir path/to/poses \
  --output_dir path/to/output \
  --model MiniMax-M2.7 \
  --api_key $MINIMAX_API_KEY \
  --base_domain https://api.minimax.io/v1/

MiniMax models supported:

• MiniMax-M2.7 — Peak Performance. Ultimate Value. Master the Complex (204K context)
• MiniMax-M2.7-highspeed — Same performance, faster and more agile (204K context)

API keys: platform.minimax.io
API reference: https://platform.minimax.io/docs/api-reference/text-openai-api