Dataset set_file_cache has no effect

[albertz]

It is being called (after handle_cached_files_in_config, after we created a dataset with cached files). But then, the _file_cache attribute is never used anywhere, as far as I can see. This doesn't seem right?

But maybe it's also not really critical?

Does it even work properly, or makes sense, when the dataset is mostly used in a subproc (via PyTorch data loader)?

If it is not really needed, let's clean this up.

(cc @NeoLegends)