Scaling Autoregressive Multi-Modal Models: Pretraining and Instruction Tuning