denoising with sentence permutation, and language sampling

[thomas-happify]

🚀 Feature request

Motivation

When training or fine tuning models, data collator provided in huggingface isn't enough.
For example, if we want to further pretrain mBART or XLM-R, where language sampling or sentence permutation are needed, which is hard to do with huggingface datasets API since it loads all language datasets at first.

Thanks!