Print Metric Issue / Feature Request

[sfc-gh-fjiang]

When I am using ArcticTraining, I got confused by the iteration summary

1. It is confusing on the total step/iter number, where currently it is self.config.epochs * len(self.train_dataloader) but when using with self.config.gradient_accumulation_steps, it will scale up by the factor. For example, if using self.config.gradient_accumulation_steps=4 and 100 global step/iters, it will always be xx/400, which is confusing. I wonder any special reason for current design.

https://github.com/snowflakedb/ArcticTraining/blob/main/arctic_training/trainer/trainer.py#L346-L354

2. It is better to add an remaining training ETA, which makes it more friendly for long training run.