Embed image and audio files in save_to_disk

[lhoestq]

Following #4184, currently a dataset saved using save_to_disk doesn't actually contain the bytes of the image or audio files. Instead it stores the path to your local files.

Adding embed_external_files and set it to True by default to save_to_disk would be kind of a breaking change since some users will get bigger Arrow files when updating the lib, but the advantages are nice:

• the resulting dataset is self contained, in case you want to delete your cache for example or share it with someone else
• users also upload these Arrow files to cloud storage via the fs parameter, and in this case they would expect to upload a self-contained dataset
• consistency with push_to_hub

This can be implemented at the same time as sharding for save_to_disk for efficiency, and reuse the helpers from push_to_hub to embed the external files.

cc @mariosasko