[interleave_dataset] sample batches from a single source at a time

### Feature request

interleave_dataset and [RandomlyCyclingMultiSourcesExamplesIterable](https://github.com/huggingface/datasets/blob/3813ce846e52824b38e53895810682f0a496a2e3/src/datasets/iterable_dataset.py#L816) enable us to sample data examples from different sources. But can we also sample batches in a similar manner (each batch only contains data from a single source)?

### Motivation

Some recent research [[1](https://blog.salesforceairesearch.com/sfr-embedded-mistral/), [2](https://arxiv.org/pdf/2310.07554)] shows that source homogenous batching can be helpful for contrastive learning. Can we add a function called `RandomlyCyclingMultiSourcesBatchesIterable` to support this functionality?

### Your contribution

I can contribute a PR. But I wonder what the best way is to test its correctness and robustness.