PyTorch DataLoader, use num_workers=1 by default?

[albertz]

Pro num_workers=1:

• It should always be faster, usually fix low computation time (PyTorch dataset use multiprocessing #1383, PyTorch with num_workers=0 and MultiProcDataset has low computation time #1435).
• It should just work.

Contra, i.e. leave the current default, num_workers=0:

• It is different from the PyTorch default.
• There might be rare cases where it does not work? (But I don't know any. I think all RETURNN datasets should work just fine.)
• Debugging might be more difficult.
• It's more complex logic.
• It adds some memory overhead.