Scaling Spoken Language Models with Syllabic Speech Tokenization