Initial recipe with AMI data

[desh2608]

This PR sets up the training and decoding pipeline with Lhotse for data preparation, PyTorch for nnet training, and k2 for graph-based decoding. AMI train set is used. Currently the per-frame accuracy is poor.

              precision    recall  f1-score   support

     silence       0.89      0.93      0.91   4578358
      speech       0.15      0.10      0.12    574642

    accuracy                           0.84   5153000
   macro avg       0.52      0.51      0.51   5153000
weighted avg       0.81      0.84      0.82   5153000