* Init commit

* Make executor use ProgramDescBind

* Change Attribute from BlockDesc to BlockDescBind

* Since we will get the program desc in RNN, just BlockDesc is not
  enough.

* add sparse support for sum op

* typo fix

* fix gpu build error

* fix unittest error

* typo fix

* infer var type and shape in op_test

* follow comments

* fix build error

* bypass some unittests depend on NetOp

* Change ProgramDesc not a global variable

* Polish code style

* Correct implement BlockDesc destructor

* Unify program as parameter name

* init parameter base class

* optimize the Comments of optimizer

* basic implimentation of optimizer

* add test_optimizer

* add no_grad_set to interface

* update optimizer.py

* python code can run

* fix some problem

* add sync_with_cpp to Python Program and Block

* sync vars and ops in block from cpp

* optimize code and add some comment

* add more check for sync

* update optimizer with return value of Backward

* rm unused code

* infer shape when create gradient vairiable

* update test_optimizer

* update test_program.py

* update backward test

* follow comment

* add target to Backward, generate var in block when call backward

* modify backward_test

* fix executor_test

* set var desc default type to LOD_TENSOR

* update backward_test

* insert loss in the top level of backward

* create grad vars for all blocks in current program

* optimize code

* update test_program.py

* only create var for newly create blocks when backward

* Add no_grad_vars for grad_op_maker

* Add unittest

* Fix unittest

* Fix unittest

* Follow comment

* Remove attr_checker from CreateGradOpDescs()

* Fix merge error

* Fix bug in backward_test.cc

Fix cycle dependencies, Fix #3583.

Fix #3582

* Also simplify pybind implementation by using OperatorBase as holder
  type.

* No POD instance is forbidden in global scope.

See
https://google.github.io/styleguide/cppguide.html#Static_and_Global_Variables

* Although backward_test/rnn_test is not pass, just comment them.