* conv cudnn 3d

* update test case

* update

* update

* follow comments and remove groups from helper

* update

* refine

* update

* follow comments2

* update

* fix compile

* "make global tensor function independently"

* "replace functor"

* "fix inline template error"

* "fix tensor array with CopyFrom"

* "fix other case use CopyFrom"

* "move the op interface hardly"

* "fix operators"

* "fix typo"

* "delete dynamic recurrent rnn and fix gru_unit in debugmode"

* "fix unique_ptr copy"

* "fix cuda copy"

* "fix namespace error"

* "removed nccl python test"

* "fix include error"

* "fix typo"

* fix copy util test

* adam sparse support

* fix gpu build error

* fix ci

* fix ci

* fix adagrad sparse update bug

* fix gpu build error

* init trieconcat_op

* add basic implementation

* add test

* add more test

* update unit test

* add PackAllSteps test

* fix PackAllSteps

* all test passed

* clean code

* remove state inside helper

* rename prob to score

* optimize RemoveFromEnd

* use deconstructor to delete BeamNode recursively

* optimize interface

* add comment to interface

* optimizer data structure

* use template to define the type of score

* use template parameter for BeamHelper

* change father to parent

* rename TrieConcat to BeamSearchOutConcat

* use LoDTensorArray

* rename BeamSearchOutConcat to BeamSearchDecode

* refine code

* remain all candidate sentence in beam_search_decode_op, do not consider endid

* use unique_ptr

* fix compare bug

* fix lod compile problem

* Add LoDRankTable

LoD Rank Table stores the `level` of `lod` which is ordered by sequence
length in descending order. It is useful when implement dynamic RNN and
is shared by dynamic RNN memory, dynamic RNN slice input and dynamic
RNN slice output operators.

* Add skeleton for array_to_lod_tensor and lod_tensor_to_array

* Add VarType::LoDTensorArray

* Add PyBind of LoDTensorArray

* Add InferVarType

* Add first unittest

* Add ut

* Add unittest

* Add unittest

* Add unittests

* update

* init

* add infershape for lod_tensor_to_array_op

* compelete array_to_lod_tensor_op

* copy data

* clean code

* clean code

* Fix unittest data

* fix bugs

* fix compile error

* Refine TensorToArrayOp

* refactor array_to_lod_tensor

* Unittest

* fix bugs

* Fix unittest

* Fix unittest

* debug

* Debug

* Fix unittest

* clean code

* refactor

* use ostream

* update test

* fix gpu build error

* make gpu test pass

* Compare Operator

* Follow comments

* Use stable_sort in lod_rank_table

It is easy to debug and test when use `stable_sort`and the time
complexity is not changed.

* Add LoDTensorArray

* Stash

* Better debug message for IsInitialized

* Stash

* Better debug message for IsInitialized

* Complete array read/write op unittests

* Add LoDRankTable

LoD Rank Table stores the `level` of `lod` which is ordered by sequence
length in descending order. It is useful when implement dynamic RNN and
is shared by dynamic RNN memory, dynamic RNN slice input and dynamic
RNN slice output operators.

* Add InferVarType

* 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 DeviceContext to Executor API

* Rewrite RNN

* Pass Python

* AddBiasOp does not care num_flatten_dims

* Stash

* Fix MacOS Compile

* Pass RNN forward

* add python test

* refactor test

* Make compile pass

* add gradopmaker

* First draft done

* Polish code

* add grad op maker and grad infershape

* Polish code

* Fix backward.cc bug

* Fix infershape

* Rename function

* add backward test

* simplify recurrent test

* Update

* Pass unittest

* Add comments & refine test

* Add comments

* refactor test

* Complete Unittest

* fix StepScopes enforce

* Remove unused unittest

* no type error

* Update

* Make RNN Pass unittest

* 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

* Init

* Stash

* Polish SaveLoadOp

* Fix CI

* Polish code

* Save GPU Tensor

* Stash

* Fix CI

* "add model format design doc"

* "add restore function"

* "add parse protobuf"

* "move necessary information to saver.proto"

* "format code"

* "add gpu option"

* "add lod info"

* "add saveop python test wrapper"

* "checkpoint reuse save operator"

* "rewrite model format design doc"

* "async support needed"

* "fix run once"

* "fix doc based on comments"

* "refine based on comments"

* "fix based comments"

* "remove persistable flag from framework.proto"

* "add IndicateDataType to restore op"

* "add save test"

* "modify save restore code"

* "modified the restore logic"

* rm checkpoint_op.cc

* rm test_checkpoint_op.py

* "get inputs outputs name from execution context"

* Saving each variable to a independent file

* Fix bugs

* Rewrite save_restore_op_test with new Python framework

* Move `SaveOp` and `RestoreOp` from OpWithKernel to OpBase

* Refine unit test of SaveOp and RestoreOp

* fix compile errorwq