* support sparse gradients
test=develop

Test PaddingRNN on V100 GPU device.

Test configuration: large model, padding mode (which is the mode using recurrentOp), one GPU.
                   
GPU memory (MiB):   6414 (this PR)     vs   6837 (without this PR)
Speed (steps/s):         10.28 (this PR)    vs    9.89 (without this PR)
 

* feature/auto_growth_allocator, test=develop

* add unittest of AlignedAllocator, test=develop

* try to turn on auto_growth to test on CI, test=develop

* fix segmentation fault in mixed_vector.h, test=develop

* add unittests, test=develop

* remove async executor and add data_feed.proto to the deps of train demo

test=develop

* not use transferscope cache in cpu case

test=develop

* adjust variable name and add comments

test=develop

* use correct format for class member in operator.h

* use correct format for class member in operator.cc

test=develop

* fix int64_t

* update fill constant op unittest

* add empty line

* feature/buffer_shared_inplace, test=develop

* refine code, test=develop

* fix elementwise_add op cpu inplace and sum inplace bug, test=develop

* add unittest and debug log, test=develop

* fix parallel_executor scope bug, polish code, test=develop

* fix sum op, activation op, single_in_place_inference bug, test=develop

* remove kLocalExecScopeName, test=develop

* fix unittest,test=develop

* fix out_var first version bug, test=develop

* follow comments,test=develop

* clean code of dim and place, test=develop

* fix failed unittests, test=develop

* test=develop, fix docker with paddle nccl problem

* test=develop, fix/gcc_4.8_ubt_link_error

* test=develop, fix code format

* Fix Mask rcnn predictor
    1. refine memory optim algorithm to support the model with the block op.
    2. output diff : modify the affine channel fuse
    3. add condition_block_infer op
add interface for setting trt calib table dir
test=develop

* add the missing files.
test=develop

test=develop

test=develop

1. Since allreduce op has 4 reduce types, We split these four reduce types into four ops
2. We also refined the collective op code, e.g. we separated the collective op kernel into CPUKernel and CUDAKernel, and remove the device specified DeviceContext parameter in template as we already knew the target DeviceContext
3. We remove the newly added Collective op role to reduce the complexity of program and graph analysis

* Int8: Fix Pooling output scale

test=develop

* Update scales quantization for certain operators

These include: concat, transpose, pool and reshape. test=develop

* Move concat minimum scale finding to quantizer

test=develop

fix data feed ptr runtime error, pipeline trainer will core in some cases, so set it nullptr as default value.

* update pe reduce config
test=develop

*  drop the local_exe_scopes of the previous parallel_executor
test=develop

* add is_runnning in communicator, test=develop

* fix prepare context redundant code problem, optimize executor by caching create_varaiables
test=develop

* supports collective training in executor

* make fetch_list runable with variables, add more unittest for use_program_cache
test=develop

* fix comment
test=develop

* use unique name for nccl_id

* supports output to stream in program_to_code

* insert sync_comm_stream before regularization; add skip_op_callstack capability in program_to_code

* set op role in collective training

* add collective op role

* remove orig file

* add build optimizer by strategy

* add collective strategy

* refine collective strategy

* add multi-process role maker

* refine strategy building factory so that we can easily plugin more strategy

* scale loss grad in collective sgd transpiler

* add support for distributed fc

* code format

* revert some features for dist fc

* add support for distributed fc training

* fix prepare context redundant code problem, optimize executor by caching create_varaiables
test=develop

* supports collective training in executor

* make fetch_list runable with variables, add more unittest for use_program_cache
test=develop

* use unique name for nccl_id

* supports output to stream in program_to_code

* insert sync_comm_stream before regularization; add skip_op_callstack capability in program_to_code

* set op role in collective training

* add collective op role

* fix comment
test=develop

* remove orig file

* add build optimizer by strategy

* add collective strategy

* refine collective strategy

* add multi-process role maker

* refine strategy building factory so that we can easily plugin more strategy

* scale loss grad in collective sgd transpiler

* add support for distributed fc

* code format

* revert some features for dist fc

* add support for distributed fc training

* test=develop
add collective op unittest standard

* test=develop
remove the test_collective directory

* test=develop
remove the test_collective directory

* remove slicegather test

* code format for reducescatter

* update attr of shard_index_op

* Modify macro nccl_helper

* remove test without distribute

* macro collective_helper

* marcro update

* test=develop
update support python3.5

* test=develop change gpu memory use to 0.1 when test

* test=develop
update ut equal func

* test=develop
set flags to 1.5

* test=develop fix pickle dumple  py35

* test=develop
fix divide in slice and add sync_comm_stream
update atol and rtol to 1e-05
rm shard_index op and test
modify read input from file to read from memory
remove origin_program in framework and add i/o in c_sync_calc_stream

* test=develop update unittest sync operator I/O

add prior_box quantization code

add scale algo rules for prior box

test=develop

test=develop

* clean build_strategy
test=develop

* DataBalanceOpHandle has been removed
test=develop

* debug

* update build_strategy.
test=develop

test=develop

(1) use channel instead of vector/BlockingQueue in Dataset，to keep same with existing implementation, and make code more readable and flexible (dataset single output channel or multi output channel). one previous memory out of limit problem is cause by not release memory after training.
(2) add Record because MultiSlotType costs too much memory (80B)，fix memory out of limit problem.
(3) add Channel, Archive in paddle/fluid/framework
(4) change dataset from shared_ptr to unique_ptr in pybind
(5) move create/destroy readers from trainer to dataset
(6) move shuffle from datafeed to dataset. dataset holds memory, datafeed is only for load data and feed data to network.
(7) fix thread num bug of Dataset when filelist size < thread num
(8) support set_queue_num in InMemoryDataset

* update execution_strategy option default value
test=develop

* fix doc error
test=develop

* remove nccl dep when the number of GPU is 1
test=develop

* fix code bug
test=develop

* update CPU_NUM config
test=develop

* update anakin-engine interfaces for content-dnn

test=develop

* support only-gpu mode of Anakin

modify eltwise parse

test=develop

* modification for thread-safe

test=develop

* Integrated template instance

test=develop

* increase template parameters

test=develop

* support MLU predictor

test=develop

* update anakin cmake files

test=develop

* update TargetWrapper::set_device

* update the initialization of anakin subgraph

test=develop

* use the default constructor of base class

test=develop

Add Pipeline Concurrency Train Mode:
- Cpp: pipeline_trainer & section_worker
- Python: PipelineOptimizer
- Add a new data_feed type: PrivateInstantDataFeed
- Add a test demo of pipeline trainer and the test model is gnn
- Do not support win32 now

* remove attribute in Allocator::Allocate, test=develop

* fix travis ci error, test=develop