Pretraining Generative Flow Networks with Inexpensive Rewards for Molecular Graph Generation