Gold-Switch: Training-Free Superposition of Slow- and Fast- Thinking LLMs