Labrador: Exploring the Limits of Masked Language Modeling for Laboratory Data