apax node: provide data path without copying the files

[PythonFZ]

Allow training an apax model without copying all the data but use the file provided.

[M-R-Schaefer]

I agree that this would be a good change. We can't do this with a simple file path though, right? so we would need to supply the train / val Atoms to apax_run directly.

[PythonFZ]

Yes, probably related to https://github.com/zincware/ZnTrack/pull/875 but it could be just train_file: Path|str = zntrack.deps_path()