Add MiniMax external LLM API provider and model paths

[octo-patch]

Reason: The agent library exposed an executable external LLM API tool but had no MiniMax provider or MiniMax-M3 / MiniMax-M2.7 model path.

What changed

• Added minimax_api(prompt, history, ...) in api.py, a new executable LLM backend that mirrors the existing mistral_api tool signature so it drops directly into an Agent toolchain ("responses, history = minimax_api(prompt, history)").
  • Talks to the MiniMax OpenAI-compatible chat completions endpoint and parses the standard choices[0].message.content response.
  • Exposes both model paths via api_model (MiniMax-M3 as the default, MiniMax-M2.7 also selectable), tracked in a MINIMAX_MODELS constant.
  • Supports both regional base URLs via region (global_en -> https://api.minimax.io/v1, cn_zh -> https://api.minimaxi.com/v1) in MINIMAX_BASE_URLS.
  • Reads its key from the MINIMAX_API_KEY environment variable; multi-turn state is threaded through history just like the existing backend.
• Imported minimax_api into phi_3_vision_mlx.py so the toolchain parser (which resolves tool names in that module namespace) can reference it.
• Added toolchain hotswap examples to examples.py covering both the default model and explicit model selection.

Uses the already-declared requests dependency; no new dependencies were added.

Checks

• python3 -m py_compile api.py examples.py
• Exercised minimax_api with a stubbed HTTP layer: verified endpoint selection per region, model routing for both models, multi-turn history accumulation, and response parsing (all assertions passed).

Note: the full MLX test suite (test.py) targets Apple silicon / mlx and downloads models, so it was not run in this environment.

[JosefAlbers]

Thanks for the PR! Before I can merge this, could you run the updated code with valid API credentials and post the output here? I can't easily verify it locally since it requires API keys. Even a short terminal log or screenshot showing the feature working would be enough.