Fix load_from_disk temporary directory

[lhoestq]

load_from_disk uses tempfile.TemporaryDirectory() instead of our get_temporary_cache_files_directory() function. This can cause the temporary directory to be deleted before the dataset object is garbage collected.

In practice, it prevents anyone from using methods like shuffle on a dataset loaded this way, because it can't write the shuffled indices in a directory that doesn't exist anymore.

In this PR I switch to using get_temporary_cache_files_directory() and I update the tests.

cc @mariosasko since you worked on get_temporary_cache_files_directory()