[Bug]: SDK sends unsupported reasoning_effort to MiniMax M3

Bug Description

The SDK sends the default reasoning_effort=high to minimax/MiniMax-M3, and LiteLLM rejects it with UnsupportedParamsError: minimax does not support parameters: ['reasoning_effort']. (Same bug as #4934, whose linked PR #4935 was closed unmerged; the faulty inference is still on main.)

Actual Behavior

get_features treats generic supports_reasoning: true metadata as proof of reasoning_effort support:

pytest tests/sdk/llm/test_model_features.py -q -k generic_reasoning
# fails on main: MiniMax M3-style metadata resolves supports_reasoning_effort=True

and select_chat_options then emits the parameter:

pytest tests/sdk/llm/test_chat_options.py -q -k minimax
# fails on main: reasoning_effort present in chat options for MiniMax M3

Expected Behavior

Generic supports_reasoning alone never enables reasoning_effort. Only an explicit override, an exact supports_reasoning_effort metadata flag, an explicit False on generic reasoning (disables), or LiteLLM-supported-params/allowlist evidence decides.

Acceptance Criteria

• MiniMax M3-style metadata (supports_reasoning: true, no exact flag) resolves supports_reasoning_effort=False.
• select_chat_options omits reasoning_effort for such models.
• Explicit overrides and exact metadata flags keep precedence; supports_reasoning: false still disables.
• Existing test_model_features.py / test_chat_options.py expectations hold.

Start with get_features and select_chat_options, then run the named focused tests: tests/sdk/llm/test_model_features.py -k generic_reasoning and tests/sdk/llm/test_chat_options.py -k minimax. Verify that generic supports_reasoning metadata does not enable reasoning_effort, while explicit overrides, exact metadata, and explicit false retain their stated precedence and the focused tests pass.