stable version

Improvements:

1. Fix eval and predict module bugs
2. Change data flow to reduce memory usage (move data transformation process into iterators)

Bugs need to be fixed:

1. When training sequence labeling model in N-fold mode, the model performance drop strangely