Handle thinking models that return wrapped content even when
Thinking=false is configured. This is a fallback strategy for models
like MiniMax-M2.1 that do not fully support the Thinking parameter.
The thinking tags and their enclosed content are completely stripped.