Add MiniMax M3 and M2.7 model metadata

[octo-patch]

Description

Add MiniMax-M3 and MiniMax-M2.7 metadata to GraphRAG's model cost registry. The metadata covers context windows, supported modalities, reasoning, prompt caching, and current token pricing, including MiniMax-M3 long-context and priority tiers.

This replaces #2428 with a clean branch based on the latest main and complete target coverage.

Proposed Changes

• Register both model IDs for OpenAI-compatible and Anthropic-compatible configurations.
• Calculate cached, long-context, and priority token costs from response usage.
• Document the global and China base URLs for both supported protocols.
• Add focused regression tests and a patch release note.

Validation

• uv run poe check
• uv run pytest tests/unit/model_cost_registry/test_model_cost_registry.py tests/integration/language_model/test_factory.py -q (12 passed)
• uv run poe test_unit (310 passed)
• uv run poe build
• uv run poe build_docs
• Captured requests through the GraphRAG adapter for both models across all four documented endpoint configurations.

Checklist

• I have tested these changes locally.
• I have reviewed the code changes.
• I have updated the documentation.
• I have added appropriate unit tests.