Dataloader perf & modularize `ctc_crf`

[maxwellzh]

Two main changes:

1. Save the file ptr such that we can avoid repeatedly open/close files in training. This is critical for dataloader perf.
2. Modularize the ctc_crf module. Now it works fully like a python module. Maybe we can further rewrite the crf init/release as a context manager in the future, with which the codes could be like

   ...
   with CRF('den_lm.fst', gpu):
   manager.run(...)
   ...

All changes have been tested on WSJ.