The Price of Prompting: Profiling Energy Use in Large Language Models Inference