Masking as an Efficient Alternative to Finetuning for Pretrained Language Models