neural machine translation model support beam search with while op

* refine conv2d_transpose layer doc

* fix conv2d_transpose doc

* fix doc

2. Fix evaluator using 'reduce_sum' op instead of 'mean' op

* also div num_samples when return cost of nce_op

And write an unittest for it

* "fix decode bug"

* "follow commnet"

* "fix error"

* "fix hook bug"

* fix based comment

* fix copyright

* fix based on comment

* registery is only used by layers
* Rename it to layer_function_generator

2. Add edit distance evaluator to evaluator.py

Users can use `a+b`, `a*10`.