* speedup split_op
test=develop

* speedup split_op
test=develop

* rename ConcatGrad to Split

* refine concat and split
test=develop

* fix compile error

test=develop

test=develop

* add detect and end2end code

* update the scale for coodinates restore

* fix merge bug with dev.

* fix merge bug with dev.

* test=develop

* fix code style test=develop

* fix code style test=develop

* test=develop

* test=develop

* test=develop

Polish styles in generate_proposals_op.

1. inline lambda functions rathar than use std::function to save var.
2. add `static inline` to template functions .cc
   * Make them static to prevent generating symbols.
   * Make them inline to give compiler a hit inline them as possible.
   * Not if the function is not static, they cannot be inlined since the
     symbols should be exported.
3. add `static` to global functions in .cc
   * Make them static to prevent generating symbols.
4. Use Vector<uint64> instead manually manange storage between devices.
5. Prefer to use platform::ForRange, so we can optimize `ForRange` by
   just changing `for_range.h` if it is needed.
6. Do not change shape of inputs

test=develop

Use OO style to rewrite memory allocation.

* Add CUDA implementation for generate_proposals_op.
* Clean code.
* Update code.

* Add roi perspective transform.

* Add roi_perspective_transform_op.

* Fix code style.

* Add python api and fix doc.

* Fix API.spec

* Fix python api.

* Fix API.spec

* Move src to detection.

* Enhence generate_proposal_labels_op
* Fix bug in generate_proposals_op
* Refine rpn_target_assign_op.
* by Bu Xingyuan, Wang Guanzhong and Dang Qingqing

* Fix bug in generate_proposals_op.
* Fix data type for RoIs.
* Refine and fix rpn_target_assign_op.
* Add the missing file bbox_util.h
* Rename BoxEncoder to BoxToDelta

* remove template operator

* "fix compile"

* "fix ci"

* "fix ci"

* Add generate_proposal_labels for Faster-RCNN.

* Add proposals generation operator for Faster-RCNN.

* Add flag to set prior_box output order.

* Add region proposal network (RPN) target assign operator and Python API for Faster-RCNN.

* Add anchor generator operator for Faster-RCNN.
* Add unittest testing.
* Add Python API.

* optimize bipartite_match_op in large scale input

* Update some doc about layers' API.

* Fix format.

* Fix example bug in random_data_generator.

* Fix example bug in dropout.

* Follow comments and some small fix for some examples.

* add normalize switch to box_coder_op

* add default prior box var

* update according to the review

* Add quad transform.

* Fix some syntax error.

* Fix CUDA kernel launch configure.

* Generalize geometry channels.

* Rename QuadTransform to PolygonRestore.

* Rename op.

* Rename op and fix computation.

* Modify CMakeLists.txt for box_restore op.

* Refine code:
1. rename op
2. uncomment unitest on GPU