One Model for One Graph: A New Perspective for Pretraining with Cross-domain Graphs