Why compute prediction over decoder_inputs instead of encoder_inputs?

Hi,

Thanks for sharing the code. I have a doubt in _pointer.py_. In the below code fragment:
```python
if feed_prev and i > 0:
inp = tf.pack(decoder_inputs)
inp = tf.transpose(inp, perm=[1, 0, 2])
inp = tf.reshape(inp, [-1, attn_length, input_size])
inp = tf.reduce_sum(inp * tf.reshape(tf.nn.softmax(output), [-1, attn_length, 1]), 1)
inp = tf.stop_gradient(inp)
inps.append(inp)
```
Here you are computing inp from decoder_inputs, but at the test time, you wouldn't really have decoder_inputs. Shouldn't it be computed from encoder_inputs rather by indexing into them?

Start by reading the feed_prev branch in _pointer.py_ and trace how decoder_inputs and encoder_inputs are supplied during training and test-time execution. Determine whether using decoder_inputs is intentional or a defect, then document the conclusion or update the implementation so the test-time input path is correct.

Written by the indexing model from the issue text.