fix some bugs/issues of the training scripts

[qindazhu]

No description provided.

[qindazhu]

There is a warning here with the latest lhotse

UserWarning: ProcessPoolExecutor using a "fork" multiprocessing detected. In some circumstances this can crash or hang the program, e.g. when using  data augmentation libsox wrappers like WavAugment or torchaudio. We suggest passing an extra argument to initialize the executor: ProcessPoolExecutor(..., mp_context=multiprocessing.get_context("spawn"))
  warnings.warn('ProcessPoolExecutor using a "fork" multiprocessing detected. '

I believe it can be fixed with @pzelasko's PR https://github.com/k2-fsa/snowfall/pull/5/files

[danpovey]

Thanks! Merging.

[qindazhu]

The scripts can now fully run with latest k2 and lhotse, but still have some issues:

1. Sometimes start_frame + num_frames > T in K2SpeechRecognitionIterableDataset lhotse-speech/lhotse#125 causes we cannot run many batches successfully.
2. The modle output is diverging, may because learning rate or model struct, we definitely need to look into this.
3. Sometimes we may get Cuda out of memory, but because 1, we can not repro it.