Further speedup DS2 training by a) turning on FLAG_rnn_use_batch 2) improve data pipeline

Improve xmap_reader_mp by adding a flush thread.

Add multi-process version of xmap_reader to take place of multi-thread version for a significant acceleration and add seqbin data parser for internal 1w data training.

fix a bug in printing parameter gradients during training for GNR.

Add preprocess and evaluation scripts for GNR.

Add seqbin data parser to adapt to internal 1w data training.

Fix the core dump bug of DS2's training in docker

Convert decoding results to unicode in DS2

Bug fix.

Fix bugs for demo_server.py.

small code cleans.

Fix bugs in preparing data scripts of DS2

Make sh scripts more robust

Add a deep stacked LSTM seq2seq model for text generation.

Avoid repeated infer for same batch in the tuning of DS2

Add a reading comprehension model.

Improve params tuning strategy for CTC beam search decoder