Patton: Language Model Pretraining on Text-Rich Networks