Masked Language Modeling for Proteins via Linearly Scalable Long-Context Transformers