* migrate load kernel

* remove load op

* fix test failed

* add save kernel

* add save_sr_kernel

* remove original save_op

* add save gpu kernel

* remove combine kernel

* add port.h include

* add save selected rows test

* remove useless kernel.h

* Add save_combine_op

* Add load_combine_op and test

* Add unit-test

* Add a delete to free buffer memory

* Add new variant of load/save

* Fix unit-test

* Add another unit test for compatibility with original save/load

* Address review comments and simplify logic

* Address review comments and simplify code - part 2

* Fix naming issues and CMake problems

* Address review comments

* Fix LoD information in tests

* Address review comments: round 2

* "change operator interface"

* "move devicepool to device_context"

* "fix operator test"

* "fix op_registry Run interface"

* "net op passed. Need to fix nccl multi-Context"

* "add nccl group function"

* "add nccl group function"

* "fix gpu count exceed 32 error"

* "fix recurrent op, nccl op"

* "change the other operators interface with Place"

* "fix typo"

* "fix pybind"

* "fix device in python side"

* "fix pybind failed"

* "add init for test"

* "fix CI"

Use posix_memalign to allocate aligned memory, since some SIMD instructions require the alignment of memory accesses.

* Init

* Stash

* Polish SaveLoadOp

* Fix CI

* Polish code

* Save GPU Tensor

* Stash

* Fix CI