Weight-sparse transformers have interpretable circuits