test=develop

* fix GetTensorFromVar
test=release/1.1

* refine GetTensorFromVar
test=develop

It seems that the blocking queue might be destroyed early than Run
method complete. It might because the Run method throw some unhandled
exception. However, it should be shared_ptr when multthread access an
resource. So change BlockingQueue as a shared_ptr.

test=develop

remove the optimized but buggy implementation

test=develop

* feat(platform): lazy initialization of devicecontext in pool

Use std::async(deferer, []{...}) to lazy initialize DeviceContext in Pool

test=develop

* Add future includes

test=develop

test=develop

* fix sum_op's GetExpectedKernelType
test=develop

* fix ci fail
test=develop

test=release/1.1

* fix nccl2 lars dist support

* put lars in momentum op

* add tests lars

* fix ci

* fix cpu kernel

* soft warning

* remove lars in test_recognize_digits.py

* move to another op

* add file

* update api.spec test=develop

* update test=develop

* fix api.spec test=develop

* wip

* wip, finish grad merge ops

* wip, finish graph build

* wip test running

* work on 1 gpu

* workable version

* update

* fix tests

* fuse broadcast op

* fix compile failed

* refine

* add batch merge test mnist

* fix CI test=develop

* fix build

* use independent bn params for batch merge test=develop

* update api.spec

* follow comments and for test

* wip

* refine tests test=develop

* follow comments test=develop

* remove startup bn modify test=develop

* follow comments test=develop

* fix merge test=develop

test=develop

* fix graph num hang test=develop

* re-enable tests test=develop

* re-enable graph num check test=develop

* fix multi device pass role check test=develop

test=develop

test=develop

please fix offline

This reverts commit 0e722c5e

test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

test=develop