support wandb log in dpo

[akk-123]

1. dpo use self.log_metrics to log metric now, please support more friendly wandb log like PPOTrainer
2. dpo need load 2 model, can you support lora model, so that only load one model

[kashif]

regarding point 2 @akk-123, the reference model is used in evaluation mode and the main model can certainly be prepared for training via Peft and that should work as per the other trainers using peft

[akk-123]

@kashif thanks. I use lora to train main model, and set save_steps to save weight, I found that there saved too many things

how can I only save adapter_model.bin and adapter_config.json?