Implement layered building

As discussed with @stas00 and @lhoestq (see also here https://github.com/huggingface/datasets/issues/2481#issuecomment-859712190):

> My suggestion for this would be to have this enabled by default.
>
> Plus I don't know if there should be a dedicated issue to that is another functionality. But I propose layered building rather than all at once. That is:
>
> 1. uncompress a handful of files via a generator enough to generate one arrow file
> 2. process arrow file 1
> 3. delete all the files that went in and aren't needed anymore.
>
> rinse and repeat.
>
> 1. This way much less disc space will be required - e.g. on JZ we won't be running into inode limitation, also it'd help with the collaborative hub training project
> 2. The user doesn't need to go and manually clean up all the huge files that were left after pre-processing
> 3. It would already include deleting temp files this issue is talking about
>
> I wonder if the new streaming API would be of help, except here the streaming would be into arrow files as the destination, rather than dataloaders.