Add decoder init & verify output

[kuke]

Resolve #781

[kuke]

Currently, the output transcriptions look like

-----------  Configuration Arguments -----------
batch_size: 128
checkpoint: train_4_gpus_resume/checkpoints/deep_asr.pass_0.checkpoint/
class_num: 101
device: GPU
frame_dim: 2640
graphs: ./decoder/graph/TLG.fst
hidden_dim: 1024
infer_feature_lst: data/aishell/feature0.scp
infer_label_lst: data/aishell/label0.scp
learning_rate: 0.00016
log_prior: ./decoder/logprior
mean_var: data/aishell/global_mean_var
minimum_batch_size: 1
parallel: True
proj_dim: 512
stacked_num: 5
vocabulary: ./decoder/graph/words.txt
------------------------------------------------
utter#0 :  我在重点长花啊

utter#1 :  温州开水浇头伏

utter#2 :  出了四周和三周积习人马呢

utter#3 :  当地悟语倒一些困难得时吼

utter#4 :  加强隆务互不慕阳结合

utter#5 :  每日经济新闻记者注意到

utter#6 :  妹妹期的欧美国家啊

utter#7 :  河南省金乡实力

utter#8 :  费主页投资存在不少经风险

utter#9 :  的新建一座速滑成国

.......

[kuke]

@pengeorge How to format the output to utilize the script for error rate measuring?