Incorrect saving of checkpoints on train_xl.py

[brycegoh]

Hi, I recently tested out the training script with --checkpointing_epoch=1 and gradient_accumulation_steps=1. I have a total dataset size of 10k with batch size of 5. This means that the script should be saving a checkpoint every 2000 steps. However, the checkpoint folder created is checkpoint-4000 and is being saved when the progress bar shows 4000 steps.

[rphly]

fixed #164

[brycegoh]

Thanks bro

[rphly]

no problem @yisol @Sang-kyung @subin-kim-cv pls see