Enable MiniMax models and wire their image/video request format

[octo-patch]

Reason: The model allowlist omitted MiniMax-M3 and MiniMax-M2.7, and the multimodal request path raised for their image/video inputs, so MiniMax-M3 could not be used for video evaluation.

Changes

• src/utils/allowed_models.json: add minimax/MiniMax-M3 and minimax/MiniMax-M2.7 to the allowed model list so they can be selected as --model / --helper_model and as evaluation models.
• mllm_tools/litellm.py: wire the MiniMax image and video request format in LiteLLMWrapper.__call__. MiniMax uses the OpenAI-compatible content schema, so image inputs are sent as image_url and video inputs as video_url (both carrying the encoded data URL). Previously any model other than Gemini/GPT raised ValueError on multimodal messages, which blocked the video-evaluation path.

MiniMax-M3 accepts text, image and video inputs; MiniMax-M2.7 is text-only. Audio and any other unsupported input type still raise a clear ValueError for MiniMax.

Checks

• python -m py_compile mllm_tools/litellm.py — passes.
• jq . src/utils/allowed_models.json — valid JSON with both new entries present.

[octo-patch]

Updated MiniMax-M3 video evaluation to use the LiteLLM media path, limited MiniMax media handling to the supported model, and added regression coverage for image/video blocks and CLI selection. I ran the unit tests, Python compile checks, JSON validation, and git diff checks.