feat: add MiniMax as LLM provider (global + China)

[kapelame]

Summary

Adds MiniMax as two first-class chat providers:

• minimax → https://api.minimax.io/v1 (international)
• minimax-cn → https://api.minimaxi.com/v1 (China domestic)

Both use the OpenAI-compatible litellm provider with api_base override — same pattern as a0_venice, nebius, and ollama_cloud already in the file. models_list.endpoint_url is set so the Model Configuration plugin can list available models dynamically.

Why two entries

MiniMax ships separate global (api.minimax.io) and China (api.minimaxi.com) endpoints with separate accounts and API keys — users on one side don't have credentials for the other. Two entries lets each user pick the one they signed up for without manually editing api_base.

Temperature clamp

MiniMax's API rejects temperature <= 0 and > 1 with HTTP 400. Without a clamp, callers passing the LiteLLM/OpenAI default of 0 (deterministic mode) hit a confusing error from the upstream API. Added a small clamp in models.py:_adjust_call_args that fires when:

• the resolved provider name is minimax or minimax-cn, OR
• the api_base URL contains minimax (catches users who set up MiniMax via the generic openai provider + custom api_base), OR
• the model name contains minimax (extra safety net)

In-range values (0 < temp <= 1) are passed through unchanged.

Tests

tests/test_minimax_provider.py adds 10 tests:

• 2 verify both providers are registered with the right api_base in the YAML
• 7 cover the temperature clamp's parametrized positive + negative cases (provider-name match, api_base match, model-name match, in-range pass-through)
• 1 verifies the clamp does NOT fire for non-MiniMax providers

All 10 pass locally with the pinned requirements.txt env.

Recommended models

MiniMax-M2.7 (latest flagship) and MiniMax-M2.7-highspeed (lower-latency variant). Users select the provider in Settings, then enter the model name.

Notes

A previous bot-generated PR (#1275) added the global minimax provider only and went stale (now has merge conflicts, no maintainer review). This PR supersedes it: covers both regions, includes the temp-clamp safety, and the test file uses the same minimal-stubbing style as tests/test_model_config_api_keys.py.