outputs wrongly

    1. disable conv activation pass by default
    2. set fc_fuser'param with_relu false while mlu fc kernel does not
    support relu
    3. change fc filter shape from 2 dim to 4 dim while input dim == 4
    4. add ToFile func in mlu tensor for debug convenience
    5. enable 4-dim input in elementwise_ops
    6. add transpose2d in utility.cc

target -> subgraph -> target -> precision -> layout

test=develop

* [LITE][OPENCL] enhance elemul, fix reshape for opencl. test=develop

* comment loop test for pad2 opencl kernel. test=develop

* [LITE][PROFILE] add sequnce statics for precision profiler. 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

* Fix OpenCL kernel of exp, tanh,

* Fix layout pass for opencl backend,

* Add how to debug cl kernel. test=develop.

* fix some error when compiling with mlu-sdk1.2.5

This reverts commit 670fcc3c.

- 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

* feature: buffer sharing, test=develop

* fix the warning of buffer.h, test=develop

* test cuda only if necessary, test=develop

* [NPU] subgraph's precision register to kAny

* get the precision from the real tensor or tensor_array of base_scope in arena/framwork

* register assign and assign_value to kAny

* [MLU] support resnet50 on MLU

* [MLU] support resnet50 on MLU

* fix: fix nlp ops input and output type
* fix: add elementwise x_dims>y_dims case

* Fix memory_resuse for opencl. test=develop

* remove useless code. test=develop

* test=devellop

* add cl file into resulted lib test=develop

* test=develop

* test=develop