Seol mar théacs é seo: Scalable syntactic inductive biases for neural language models