The tokenizer for literate modes was incorretly detecting code lines in text. This PR fixes that, and allows for empty code lines.