align the dygraph behavior of simnet with static graph

[FrostML]

【WIP】In dygraph similarity_net, list problems before this fix:

• use fluid.layers.reshape(size=[-1, 1]) to guarantee that the shape of Embedding's input is [xx, xx, 1], which is NOT recommended and NOT proper for batch input.
• the behavior for lstm and gru to compute sequence representation is NOT align with static graph, which is fluid.layers.reduce_sum while sequence_last_step for static graph.

Due to pre-commit, code format has been changed. Mark changes as follow to help review.

• gru.py
  

• lstm.py
  

• paddle_layers.py
  

• run_classifier.py
  delete all fluid.layers.reshape

• utils.py
  

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}