Should You Mask 15% in Masked Language Modeling?