Add lstm with recurrent projection layer operator

enhance the machine_translation model in unittest.

Feature/add demo for parallel.do

* fix call once logic

* clean up

* further clean up

* fix root number

* fix call once logic

* clean up

* further clean up

Transpiler: fix pserver crash due to split var name check.

* Add distributed implementation for recommender system

* Addressing code review feedback

In notest_dist_label_semantic_roles.py, "emb" is matched with
"embedding_1.w_0", but they are two irrevalent vars.

Fixes: https://github.com/PaddlePaddle/Paddle/issues/7701

fix openblas build

* remove pick dependency

* fix bug

* small fix

* modify executor.py for save and load

* clean code

* Add usage example

* refine executor run function

* fix bug

* refine executor code

* fix block bug

* fix comments

* fix bug

* fix pass num

Fix get_inference_program

fix a bug

fix a bug

* aysnc send/recv, seriliaze/deserialize using threadpool

* implement paralell deserialization correctly

Add `ProgramDesc` storage

Rename is_compile_gpu to is_compiled_with_cuda

 Make get_inference_program support for Evaluator.

Add the wrapper for multi-head scaled dot product attention.

Disable the unstable unittest

Test parallel do with multiple input

Unify the recognize_digits

Merge branch 'feature/rename_is_compile_gpu_to_is_compiled_with_cuda' into feature/add_demo_for_parallel.do