FIRP: Faster LLM inference via future intermediate representation prediction