* [LITE][PROFILER] Fix unit test segfault when profiler on. test=develop

* reduce .so size. test=develop

* compile all targets when LITE_ON_TINY_PUBLISH=OFF

* unordered_map is more convenient when key is customized class

* test=develop

[LITE][OPENCL] Enhance Profiler for OpenCL with in/out/filter shape, macs/macs_ps, real backend kernel etc. (#3641)

* [LITE][OPENCL] Enhance Precision Profiler for OpenCL. test=develop

refactor(*): reduce Wsign-compare warning

lite cuda support exec multi-stream

* split precision profiler from performance profiler. test=develop

test=develop

* [LITE][OPENCL] clean code for opencl. test=develop

* [LITE][PROFILER] Enhance Precision Profiler. test=develop

* delete useless var in profiler. test=develop

* add ocl header. test=develop

- 增加cuda c++ demo.
- 考虑到检测模型尾部一般是multiclass_nms，该kernel为host，如果fetch kernel为cuda的话，则会在此处插入无用的io_copy(host->cuda)，由于该原因，注释掉fetch的cuda kernel. 默认使用host的fetch kernel. 此处暗中进行的行为：每次predictor run完，都会默认把数据从cuda拷贝到cpu

Optimize cuda kernel and remove io_copy added by default due to missing fetch_cuda kernel

fix fpga lite_tensor compile bug
     add fake quantize_abs_max op
     test=develop

Optimize the execution of RuntimeProgram by saving the bool whether the op is feed/fetch op. (#2703)

test=develop

* [ARM] change global pooling choose kernel policy, test=develop

* update profiler, test=develop

* warm up times of profiler, test=develop

[LITE][NPU][XPU] Refine subgraph pass, and support NPU/XPU model generation at execution time (#2576)

修改了选kernel的逻辑，默认从模型文件中读取出lod_tensor的data type，在static_kernel_pick pass中如果kernel输入输出的类型与读取的data type完全一致，则选择该Kernel的概率增大。

- 增加 从模型文件__model__读取lod_tensor的data type到cpp::vardesc

- program中增加unordered_map<string, type>字段，并在 Program::PrepareWorkspace中对该字段赋值

- 修改了node.h文件，将const Type* 更改为Type*，并在SSAGraph::Build过程中为符合条件的type*赋值

- static_kernel_pick_pass中添加新规则，如果kernel的输入类型输出类型与__model__中存储的类型的一致，则score*=2。

- 支持模型中用到sequence_reverse_float kernel（输入输出均为float）和sequence_reverse_int64 kernel（输入输出均为int64），能够根据输入输出type选kernel

* add arm split lod tensor, test=develop

* add arm merge lod tensor, test=develop

* update split merge lod tensor, test=develop

* add reduce_prob op, test=develop

* support mask_rcnn succeed, test=develop

* [LOG] macro for vlog. test=develop

* make inceptionv4, resnet50, googlenet can run on x86 paltform and fix the compare part in x86 unittests, test=develop

* fix googlenet tests for benchmark record, test=develop

* [framework][profile] fix profile dump bug when op is feed and fetch test=develop (sangoly)

* init: delete feed and fetch op, using zero copy
test=develop

* delete the unused test
test=develop

* add conv int8 support(in condition which the input or output channel not be the times of 4)
add add_kernel for cuda.

* can run yolov3 fp32
test=develop

* 1. fix bug with yolov3 run
test=develop

* fix: fix model parser and save bug

* style: delete debug code

* fix: fix light_predictor program run model with subblock bug

* Fix timer of arm cpu profiler. test=develop

* Fix un-added op in cmake.test=develop

* fix cmake error

* fix cmake error, test=develop

* Fix pass sequence. test=develop

* replace option with lite_option. test=develop

* disable profile mode by default. test=develop

* Fix error option name. test=develop

* Add precision and persistable attrs for the tensor. And fix cxx light and full api demo.

* update precision2string methods. test=develop

* move the save logic to the front of the run in mobilenetv1_full_api.cc, test=develop.

* add comments for UpdateVarsOfProgram. test=develop