bug caused by lazy malloc of tensor memory. while batchsize large than
1, output tensor is divided(shared memory), and each divided tensor do malloc.

add multiclass_nms2 kernel, 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

* rewrite multiclass_nms according to fluid, test=develop

* fix param load bug in box_coder and multiclass_nms op, test=develop

* fix value transfor error in multiclass_nms, test=develop

* 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