Investigating Energy Efficiency and Performance Trade-offs in LLM Inference Across Tasks and DVFS Settings