Interpretable Language Modeling via Induction-head Ngram Models