Intelligence per Watt: Measuring Intelligence Efficiency of Local AI