Measuring the Knowledge Acquisition-Utilization Gap in Pretrained Language Models