Energy Use of AI Inference: Efficiency Pathways and Test-Time Compute