SprintCacheDataset issue with torch backend

[robin-p-schmitt]

For the latest RETURNN, when using torch backend and SprintCacheDataset, I get this error.

File "/usr/local/lib/python3.10/dist-packages/torch/utils/data/datapipes/datapipe.py", line 179, in IterDataPipe.__reduce_ex__
    line: return super().__reduce_ex__(*args, **kwargs)
    locals:
      super = <builtin> <class 'super'>
      __reduce_ex__ = <not found>
      args = <local> (4,)
      kwargs = <local> {}
  File "/usr/local/lib/python3.10/dist-packages/torch/utils/data/datapipes/datapipe.py", line 354, in _DataPipeSerializationWrapper.__getstate__
    line: value = pickle.dumps(self._datapipe)
    locals:
      value = <not found>
      pickle = <global> <module 'pickle' from '/usr/lib/python3.10/pickle.py'>
      pickle.dumps = <global> <built-in function dumps>
      self = <local> _IterDataPipeSerializationWrapper
      self._datapipe = <local> BatchingIterDataPipe
TypeError: cannot pickle '_io.BufferedReader' object

The full error log and returnn config can be found here: https://gist.github.com/robin-p-schmitt/dca90f178d1bfc086f8faadb35211406