* Support Sync Batch Norm.
* Note, do not enable it in one device.

Usage:

build_strategy = fluid.BuildStrategy()
build_strategy.sync_batch_norm = True
binary = fluid.compiler.CompiledProgram(tp).with_data_parallel(
        loss_name=loss_mean.name,
        build_strategy=build_strategy)

test=develop

* wip

* wip

* wip

* wip for test

* add fp16 tests test=develop

* fix cpu build test=develop

* fix test=develop

* fix py3 tests test=develop

* fix lr_scheduler dtype test=develop

* fix test=dvelop

* test fix ci compile test=develop

* fix build and merge test=develop

* fallback momentumop change to general test=develop

* make fp16 lr schedule simple test=develop

* fix ut test=develop

* fix tests test=develop

* remove fp16 learning rate cast test=develop

This reverts commit 3d750f9c.

* wip

* wip

* wip

* wip for test

* add fp16 tests test=develop

* fix cpu build test=develop

* fix test=develop

* fix py3 tests test=develop

* fix lr_scheduler dtype test=develop

* fix test=dvelop

* test fix ci compile test=develop

* fix build and merge test=develop

* fallback momentumop change to general test=develop

test=develop

* wip multi process multi gpu dist training

* workable for p2p

* update test=develop

* change back env name test=develop

* fix alloc init

* fix cpu build test=devlop

* fix mac tests test=develop

* refine code

* refine test=develop

test=develop

Fix code to support cpplint syntax check

test=develop

* support test using executor without reader

* run imgnet

* update fluid benchmark

* wip

* update

* update all models

* support pyreader

* update

* clean up

* make profile batches contollable

* update API.spec

* update scripts

* clean dockerfile

* update

* clean comments

* add scope argument docstring

* use num_trainers to determine nccl init comms

* 1. Create buddy allocator in each places before NcclBcast the variables
2. Check the memory usage of ALL gpus rather than the first one

* 1. Make NCCLGroupGuard guards only the ncclBcast part, which avoid ncclGroupEnd blocking the exception throwing
2. NOTE the usage of NCCLGroupGuard

* Remove the memory usage check of gpus

* Fix code style

Make vec<Tensor> can be serialized to RecordIO

* Update tensor_util.h

* Update with moved TensorDesc

* Fix tensur_utils.cu

* Update

* Update

* Update

* Update

* Make tensor_util.cu a symbolic link