DataLoaderIter has been renamed to _DataLoaderIter "to discourage
subclassing" (pytorch/pytorch#5619).

The custom DelayededDataLoader mechanism is still useful for two reasons:
- Prevents crashes when KeyboardInterrupts are triggered
  "at the wrong time". Replacing everything with vanilla PyTorch
  DataLoaders leads to a complete crash when <C-c>-ing during background
  process initialization
  (RuntimeError: DataLoader worker (pid 15559) exited unexpectedly with
   exit code 1).
  But:
  When initialization is finished (so, after ~2 seconds), the vanilla
  PyTorch DataLoader does recover from interrupts: When continuing, new
  background processes are started automatically and continue making batches.
- Reduced visual noise in the terminal (large tracebacks are suppressed)

I am not sure if those reasons are good enough for keeping it though...

…rch#5619)