The Effect of Scheduling and Preemption on the Efficiency of LLM Inference Serving