Division-of-Thoughts: Harnessing Hybrid Language Model Synergy for Efficient On-Device Agents