Avoid downloading the whole dataset when only README.me has been touched on hub.

### Feature request

`datasets.load_dataset()` triggers a new download of the **whole dataset** when the README.md file has been touched on huggingface hub, even if data files / parquet files are the exact same.

I think the current behaviour of the load_dataset function is triggered whenever a change of the hash of latest commit on huggingface hub, but is there a clever way to only download again the dataset **if and only if** data is modified ?

### Motivation

The current behaviour is a waste of network bandwidth / disk space / research time.

### Your contribution

I don't have time to submit a PR, but I hope a simple solution will emerge from this issue !