Add Debug utility to be able to preview first samples used for training

[NouamaneTazi]

Ensuring that we're using the correct data is crucial for training and could avoid hours of debugging. It would be cool to have a debug utility that can pull the couple first samples from the dataloader just like we do in trainer, then detokenize them and print them to the user

[Wodlfvllf]

I wanna take this problem. Can you assign it if this thing is still required?