test=develop

* Inception fusion operator.
* Support horizontal layer fusion in conv_fusion_op.
* Search conv algo strategy for variable-length input.
   search N times and cache the searched algos. For other input, choose the algo of input whose area is closest to this input.

* wip

* wip

* refactor no.1 dir structure test=develop

* fix linking test=develop

* fix includes test=develop

* fix build test=develop

* fix build test=develop

* refine batch merge pass

* refine batch merge pass test=develop

test=develop

test=develop

test=develop

* refine parallel_exe
test=develop

* rename shared_var_device

* code refine

* add test_weight_decay

* remove Sort
test=develop

* Add SortForReduce
test=develop

* code refine
test=develop

* follow comment
test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

test=develop

add module cleanup
test=develop

test=develop

test=develop

test=develop

This reverts commit 7f6d8ace.

test=develop

test=develop

test=develop

* code style fix, test=develop

test=develop

test=develop

* Make topk op support variable k.
test=develop

* Fix tensor type.
test=develop

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

Stack OP has much higher CPU cost than expected in release mode.
Caused by DebugStringEx() in base class OperatorWithKernel. Actually
this issue occur for each OP which hasn't implement it's own
GetExpectedKernelType().

test=develop

* refine tensor
test=develop

* refine tensor
test=develop

* fix device_context log
test=develop