* exhaustive search for cuDNN conv.
* Refine code and add unit testing.
* Fix model load in fluid/inference and unit testing in conv2d
* Follow comments.
* Fix compiling test=develop

* fix input<tensor>
test=develop

* fix split_ids
test=develop

* ElementwiseMul should not support SelectedRows

* fix scale op
test=develop

* change GetTensorFromVar() method to GetTensorOrSelectedRowsFromVar()

* fix operator

* refine MultiOutput

* fix MultiOutput
test=develop

* disable test_dist_save_load
test=develop

* fix elementwise_op
test=develop

* add get_sparse_as_op
test=develop

* add info for check
test=develop

* rename get_sparse_as_op with extract_rows_as_op.
test=develop

* elementwise doesn't support selected_rows

* fix regularizer

* remove extract_rows_as
test=develop

* fix ci
test=develop

* add test for sum_op

* fix regularizer
test=develop

*  test=develop

* fix pserver weight decay multi inputs test=develop

test=develop

Dry run tests will skip `Op.Run` and just perform job scheduling. It helps to analysis dead lock in PE.

test=develop

test=develop

This reverts commit ce7d9b07.

* exhaustive search for cuDNN conv.
* Refine code and add unit testing.
* Clean code
* Fix model load in fluid/inference and unit testing in conv2d
* Follow comments.

test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

* wip

* add ref_by_trainer_id op

* ready to test

* fix ref inputs

* refine rpc_op_handle

* fix merge bug

test=develop

test=develop

test=develop

* enable pyreader use pin memory

* add py reader pin memory test test=develop

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

added depthwise conv mkldnn pass which for MKLDNN changes depthwise_conv operator to conv operator because for mkldnn this is the same api
test=develop

test=develop