feat: add MiniMax as LLM provider backend with M3 upgrade

[octo-patch]

Summary

Add MiniMax as a first-class LLM provider backend, using the OpenAI-compatible API.

Update: Upgraded default model from MiniMax-M2.7 to MiniMax-M3, the latest flagship model with 512K context, 128K max output, and image input support.

Changes

• Add minimax to ClientBackend type in rlm/core/types.py
• Route minimax backend to OpenAIClient with MiniMax base URL in rlm/clients/__init__.py
• Auto-detect MINIMAX_API_KEY environment variable in rlm/clients/openai.py
• MiniMax-M3 set as the default model; M2.7 and M2.7-highspeed retained as alternatives
• Older models (M2.5, M2.5-highspeed) removed from the model list
• 18 unit tests + 3 integration tests in tests/clients/test_minimax.py

Models

Model	Description
MiniMax-M3 (default)	Latest flagship model — 512K context, 128K max output, image input support
MiniMax-M2.7	Previous generation, kept for backwards compatibility
MiniMax-M2.7-highspeed	High-speed version of M2.7 for low-latency scenarios

Usage

from rlm.clients import get_client

client = get_client(
    backend="minimax",
    backend_kwargs={"model_name": "MiniMax-M3"},
)
result = client.completion("Hello!")

Testing

• All 21 tests passing (18 unit + 3 integration)
• Integration tested with real MiniMax API using M3, M2.7, and M2.7-highspeed