Option for non HDF5

[northanapon]

I had a few problems with HDF5 format on my dataset, so I add an additional code:

1. Added a script for preprocessing in torch.
2. Added an option to training.lua to use data from the new script
3. Added a new data loader that returns data continuously from batch to batch

[AlekzNet]

You may want to add sorting (e.g. like in char-rnn) to avoid index mismatching, when you use a pre-trained model on a different corpus:
#43
#37 (comment)