feat: upgrade MiniMax default model to M3

[octo-patch]

Summary

Upgrade MiniMax model references in tests to use MiniMax-M3 as the current default, while still asserting capability behavior for the M2.7 family that remains supported.

Changes

• tests/services/llm/test_capabilities.py — extend test_minimax_openai_compat_supports_tools_without_response_format to cover MiniMax-M3 and MiniMax-M2.7 in addition to the existing MiniMax-M2.7-highspeed assertions.
• tests/services/config/test_provider_runtime.py — update the sample MiniMax model from MiniMax-M2.7 to MiniMax-M3 in test_llm_minimax_binding_uses_minimaxi_endpoint so the test exercises the new default model.
• tests/core/test_agentic_client_provider_kwargs.py — update the minimax_anthropic native-tool-calling assertion from MiniMax-M2 to MiniMax-M3. (MiniMax-M2 is no longer a first-class target on the MiniMax side.)

Why

MiniMax-M3 is the current default model on the MiniMax platform. It introduces:

• 512K context window
• 128K max output
• Image input support on both OpenAI-compatible and Anthropic-compatible endpoints

DeepTutor's MiniMax / minimax_anthropic providers already accept arbitrary model strings, so the only meaningful in-tree change is to refresh the model identifiers used in tests so they match what users will actually configure today. The provider capability matrix (no response_format, tools enabled, etc.) is unchanged.

Test Plan

• Existing tests still cover MiniMax-M2.7-highspeed behavior.
• New assertions confirm MiniMax-M3 and MiniMax-M2.7 get the same capability flags via the existing provider entry (supports_tools=True, supports_response_format=False).
• No production code paths were touched — provider registry, capabilities map, and reasoning params remain as-is.