Created by: zhiqiu

Update ptb model to save training time.

• reduce number of reshape

• reduce number of conversion between VarBase and numpy.

Before: 117.9s / epoch on v100

After: 108.6s / epoch on v100