Created by: lcy-seso

• enhance the reorder_lod_tensor_by_rank operator. PR https://github.com/PaddlePaddle/Paddle/pull/7251
• add the external memory (the static input in old PaddlePaddle) to dynamicRNN. PR https://github.com/PaddlePaddle/Paddle/pull/7434
• make the RNN search model successfully run on Fluid.
  • https://github.com/dzhwinter/benchmark/blob/master/fluid/machine_translation.py
• test the RNN encoder-decoder without attention (This is another test case for reorder_lod_tensor_by_rank operator, so it should be tested also).
• enhance the documentation of operators used in RNN Search.
• debug and unittest the beam search related operators. @jacquesqiao is working on this, related PR (WIP) https://github.com/PaddlePaddle/Paddle/pull/7340 .
  • add lod tensor ToAbsOffset test
  • Topk share lod
  • change data type of beam_search op
  • Add get lod for debug
  • nmt model with beam search
• wrapper the beam search: issue https://github.com/PaddlePaddle/Paddle/issues/6900 .
• provide a more friendly Python wrapper which unifies training and generating sequences.