One Problem about backward mask in Language model cost function

Hi,

In the `laber.py` line 243:
```pyton
                lmcost_bw_mask = tf.sequence_mask(sentence_lengths, maxlen=tf.shape(target_ids)[1])[:,:-1]
```
The mask have some issue, for example
```
origin_seq: 1 2 3 4 0 0 0
origin_mask: 1 1 1 1 0 0 0
lmcost_bw_mask: 1 1 1 1 0 0
the correct lmcost_bw_mask here should be: 1 1 1 0 0 0
```