fix: move quant op to basic

skip fusing quantized conv2d + relu6 for now

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

add multiclass_nms2 kernel, test=develop

* fix caliberate error, values from [-128, 127] to [-127, 127], test=develop
* add classify demo and detection demo, test=develop

* add weight quant for combined model after optimized, test=develop

* fix compile. test=develop

* make full_publish to fix link bug in mask demo

add cv docs, test=document_fix

* Add reduce sum op test
test=develop

test=develop

* mobilenetv1/v2,mnasnet,yolonano,sr models passed with opencl image2d kernel. test=develop

* Fix gflag shape define. test=develop

* Replace Softsign Eigen with c implementation
test=develop

* Optimize softmax. When the input tensor is 2-D and axis is 1, there is no need to resize.

* Optimize the gru, avoid calling Tensor::Slice.
test=develop

* Remove a std::vector in softmax.
test=develop

* Define CalculateSeqWidth to get the width of a sequence.
test=develop

* comment All opencl buffer kernels. test=develop

* refactor conv, depthwise into one routing selection. test=develop

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

Update mask demo docs, add model download from paddlehub

* [LITE][BM] support VGG,Inception,Mobilenet,Darknet, test=develop

* modify opt name test=develop

* fix code style test=develop

* test=develop

* replace gru RowWiseAdd Eigen with c implementation
test=develop

* [LITE][OPENCL] Add opencl image2d conv3x3. test=develop