[Streaming] Only load requested splits without resolving files for the other splits

[lhoestq]

e.g. thangvip/cosmopedia_vi_math has 300 splits and it takes a very long time to load only one split.

This is due to load_dataset() resolving the files of all the splits even if only one is needed.

In dataset-viewer the splits are loaded in different jobs so it results in 300 jobs that resolve 300 splits -> 90k calls to /paths-info

[lhoestq]

This should help fixing this issue: #6832

[akaashdash]

I'm having a similar issue when using splices:

It seems to be downloading, loading, and generating splits using the entire dataset.