Prompt-prompted Mixture of Experts for Efficient LLM Generation