* [ARM] sgemv support transA, test=develop

* add sgemv ut, test=develop

* update lookup_table arm x86, test=develop

* add lookup_table_v2_op for compatibility, test=develop

align the lite nearest， bilinear op to fluid on arm and cuda

* update concat and unsqueeze, test=develop

* fix pooling bug and speed

* fix build error

* delete VLOG in pool, test=develop

* add openmp, test=develop

* fix lite/kernels/arm/pool_compute_test basic_pooling compute error bug, test=develop

* move multiclass_nms kernel back to host test=develop

* move layer_norm OP and arm_kernel into extra type since it's added after release/v2.0-beta1 and not related with CV test=develop

* fix code_style test=develop

mv deviceinfo.workspace and other relative member to thread_local_storage

We have changed 11 arm_kernels into extra type in #2347 , which has caused test_compiling failure. In this PR , we move their 11 related arm_kernel_test into build_extra=ON

update reshape op to support multiple input types of shape.
priority: input(ShapeTensor) > input(Shape) > attr(shape)

update slice op to support multiple iput types of starts and ends.
priority: input(StartsTensor) > input(StartsTensorList) > attr(starts)

fill_constant kernel only registered float type, only the float data type is produced, which is obviously a bug.

Now, produce data based on the data type attr.

By the way, fix the cast kernel bug.

* move some basic ops into extra type to reduce library size test=develop (#2347)

test=develop

* fix: fix fc batch bug

* fix:fix fc bug;test=develop

* feat: add beam_search_special function for support nlp model

* fix: add beam_search_compute kernel input and output

* feat: add assign op & copy_compute kernel

* feat: add fill_const_batch_size_like op & kernel

* feat: add layer_norm op and kernel and ut

* fix: fix some bugs
    fix mul_op infer_shape bug when x_dim_idx = 2, x_dims.size()=3 & y_dim_idx = 1, y_dims.size()=2
    fix elementwise_compute bug when y axis is all 1
    fix beam_search choose math_func wrong bug
    fix layer_norm get attr bug
    fix fill_constant_batch_size_like shape_set bug

* feat: add gather op and kernel & and transform ut

* feats: add ops and fix bugs to support transformer op
       fix type_cast passes to skip `while`
       fix elementwise infer_shape bug when x.dims=3 and y.dims={1} & axis=0
       fix lookup_table compute bug
       fix read_from_array/beam_search/increment/compate/gather ops data_type problems

* fix:
    transfomer ut add word read inferface
    fix copy/gather/norm/layer_norm include path problem

* fix:debug info

* fix: fix input reshape bug

* fix: fix norm bug

* style: style fix & test=develop

* style: fix operators cmakelist

* style: fix operators cmakelist; test=develop

* fix and test=develop

* fix and test=develop

* style: style fix; test=develop

* cuda add softmax and pool op

* * fix armlinux can find sys/system_properties.h
* conv add padding_algorithm
test=develop

* delete padding_algorithm in op param, test=develop

* fix bugs, test=develop

* update con_dw

* update

* add conv_depthwise_3x3s1.cc and conv_depthwise_3x3s2.cc

* add conv_depthwise_3x3s1_fp32 and conv_depthwise_3x3s2_fp32

* add new conv_dw

* only support conv_dw pad=0, 1

* add conv_dw_s1 conv_dw_s2 fp32

*     //conv2_func _impl2{nullptr};
update conv_dw, add conv_3x3s1 and conv_3x3s2, pad=[0,1]

* fix format, test=develop

* fix formmat, test=develop

* fix conv_transpose error

* fix build error, enable basic test of conv_transpose, test=develop

*  imporve prepack_input func speed in int8 3x3s1 dw conv

* fix code style

* fix code style

* improve 3x3s1 dw fp32 conv speed a little

* arm add 5x5s1 int8 dw conv, test=develop

* add cast from uint8 to float, test=develop

* fix gemm_int8, gemv-int8 and conv-int8 math function, add float bias

* change conv impl

* neon int8 kernel support float bias

* arm compute kernel support float bias

* add math_test target

* add tensor utils for testing, fix sgemm ut error

* add gemm_int8 unit test, support float bias

* fix build script

* add conv compute unit test for arm

* fix build script, test=develop

* fix fp32 dw conv3x3s1, test=develop

* add fp32 dw conv3x3s1, test=develop

* add armv7 fp32 dw conv3x3s1, test=develop

* add fp32 depthwise conv3x3s2, test=develop

* fix fp32 conv3x3 depthwise build error, test=develop

* fix gemm_like conv trans weights error, test=develop

* fix int8 depthwise conv3x3 error, test=develop

* turn on all test for arm fp32 conv, test=develop

* fix int8 conv1x1 error

* fix int8 direct conv3x3s1 error, test=develop

* fix int8 direct conv3x3s2, test=develop

* turn on all test for arm int8 conv, test=develop

* fix int8 fc error, change mobilenetv1-int8 ground-truth result to fluid, test=develop

* remove debug info, strip ut binary, test=develop

* fix conv compute error, test=develop

* change Init() to ReInitWhenNeeded(), test=develop

* fix code style, test=develop

* remote engine_test, test=develop

* fix building server tests error, test=develop

* fix sdot clang build error, test=develop

* fix sgemm ut timeout error, test=develop

* fix clang build error, test=develop

* turn off math basic test due to ci time out, test=develop

* fix conv_int8 ut error, test=develop

* modify norm kernel to run caffe_facedetection model

* reserve bind norm, remove calc of norm output

* add unsqueeze and range op. modify concat op test=develop

* modify exception in range_test_x86

* add assign_value op, arm kernel and test, add fluid_type, test=develop

* add hard_sigmoid, test=develop

* add ops for faster rcnn

* disable test for generate_proposals and roi_align, test=develop

* remove .swp file

* remove log in tensor slice

* finish the unit test for roi_align, test=develop

* add box_clip op and fix tensor slice bug

* remove add four op twice

* rewrite the implement for box_coder and sequence_expand, add faster_rcnn_test, test=develop

* fix test bug of box_clip in x86 server, test=develop

* fix elementwise_op bug when the shape of input y is 1, test=develop

* fix elementwise ops bug when the shape of input y is 1, test=develop

* add stack op and add reduce_mean op and their unit tests

* modify stack op output name and modify the for loop in reduce_mean op

* add HasAttr for slice op

* add yolo_box_compute cuda

* move multiclass_nms(arm) to host

* add lod in scale op

* add yolo_box_cuda cmake config

* modify shuffle_channel_fuse and transpose_softmax_transpose_fuse to support run ssd model. test=develop

* reshape and transpose op don't have xshape output.

* modify yolo_box_compute_cuda, use tensor to manage cuda memory test=develop

* add yolo_box use kernel test=develop

ad ops for faster rcnn, including affine_channel, anchor_generator, generate_proposals and roi_align (#1895)

* add ops for faster rcnn

* disable test for generate_proposals and roi_align, test=develop

* remove .swp file

* remove log in tensor slice

* finish the unit test for roi_align, test=develop

* modify cast op, test=develop

* modify cast op and remove warning in argmax_test, test=develop