1add roi align mlu kernel
2add LITE_BUILD_EXTRA when building test for mlu roialign kernels
3modify lite.cmake to fix compiling error when supporting mlu kernelxx.o exists

* (feat): add cnplugin cmake

* (feat): add cnplugin deps to backends/CMakeLists.txt and utility.h

* (feat): add box_coder converter

* (feat): add density_prior_box op

* (feat) add multiclass nms converter, while workspace as static tensor

* (ref): change nms test input from file to random, add iou

* (ref): add density_prior_box & box_coder to paddle_use_bridges.h

modify mlu's TargetWrapper to situation when creating and running predictor in different threads

1 add int32 mlu io copy out support;
2 do not ModifyInputOutputDataType for int32

1 flatten & reshape 之前没有放到 paddle_use_bridges 中，现在放进去；
2 reshape 中 cnmlCreateNdReshapeOpParam dim_num 之前误写成4，应该是 output_dims.size()

(feat): add idenity cast eliminate pass, only remove int32 to int32 cast for now (#110)

1 Fix CI test

2 refactor: reduce hwtime log to ease analyzing for CI
subgraph will rebuild while batchsize changable is on and encounter a
unsupported operator. graph will be destructed and recreated, in which Nan
hw perf log is printed

3 fix googlenet gtest check

* (bugfix): emplace_back may reallocate, make cast_op ptr detached

* fix: wrong op sequence will cause node in graph not added
Co-authored-by: N--get <zhaoying@cambricon.com>
Co-authored-by: Ndingminghui <dingminghui@cambricon.com>

subgraph will rebuild while batchsize changable is on and encounter a
unsupported operator. graph will be destructed and recreated, in which Nan
hw perf log is printed

before c++11: `iterator erase( iterator pos );`
after c++11: `iterator erase( const_iterator pos );`

(ref): insert trans before and after transpose