Since input tensor is modified to NHWC, which is not compatible to cast
kernel in NCHW layout, type_layout_cast_pass will insert a layout
instruction to transform data. Use NHWC cast kernel to avoid cast layout

while filter weight is quanted elementwise, use maximum of
weight_scale instead of mininum to keep quant weight in range of int8

for multiple-use args, only cast once

           testcase

* Optimize matmul for size(x_dims)=2  size(y_dims)>2

* [NPU] add shape bridge

move shape arm kernel to host

* enhance compare arm kernel

* [NPU] add gather op bridge

* enable reshape arm ut

* [NPU] add lookup_table bridge

lite cuda support exec multi-stream

* [LITE][OPENCL][Image] lazy init cl::event , test=develop

* [LITE][OPENCL][Image] lazy init cl::event , test=develop

[LITE][OPENCL][Image] use unique_ptr to wrap filter and bias in conv image compute , test=develop (#3388)

* [LITE][BM] fix reshape infer shape issue, test=develop

* [LITE][BM] with testing=on, test=develop

1. clean code;
2. change `cl::Kernel` from unique to shared ptr;
3. `reset` `cl::Program` and `clear` `device_info_` in destroyed of CLRuntime;
4. remove clFlush in destroyed of CLRuntime.

fix opencl hang on mali

* Optimize weight quantizaion, test=develop

* update profile info(add new element), test=develop