Fix bug in DNADataset that occurs on newer pytorch versions
I installed PyTorch 1.10 and started getting an error because of the base class ordering of DNADataset
This is because for some reason the Dataset base class grew an
(undocumented AFAICT) system for registering new methods on an existing
Dataset class through a new "function" registry. It's probably used for
something internally but I'm not sure. It works by implementing
Dataset.__getattr__
to look up registered functions in the function
registry.
However, it does not behave as a good citizen, and does not call super().getattr(), and instead raises AttributeError on anything not found in its function registry (I will open an upstream bug about this).
Therefore ConfigMixin.getattr was not being called unless it is placed before Dataset in the MRO.