Set the Clock: Temporal Alignment of Pretrained Language Models