* fp16 result ok

* change -DWITH_NVINFER_PLUGIN toconfig.EnableTensorRtOSS

* auto detect special slice op converter for ernie with trt oss

* ernie oss only support fp16

* fix special_slice_plugin serialize bug

* matmul in tensorrt ok

* ernie unittest ok

* add matmul tensorrt unittest

* remove demo code

* pool2d cuda kernel supports fp16

* fix compile issue of template

* add ut

* Add rnn_op.
test=develop

* Fix rnn_op grad maker's drop_empty_grad.
test=develop

This PR is follow up of #28213. On that PR we tried to decrease GPU usage, however the CI still randomly failed. So I added retry logic for the initialization of nccl and cusolver. If the initialization failed, we can retry to avoid the random failure.

* add generate_proposals_v2 op

* update version pybind to v2.4.3, test=develop

* update unittests, test=develop

* enable infer model running CAPI

* output size should bigger than 0

* add op_function_generator for box_coder

* fix format

* 01:Modify the shell script according to the specification

* 01:Modify the shell script according to the specification

* Enhance multiclass_nms to support LoD for dygraph mode

* fix some error in multiclass_nms

* update GetLodFromRoisNum to GetNmsLodFromRoisNum

test=develop

* update matrix nms op to api 2.0

* modify code according to review

* add fuse_bn_add_act pass

* add multiple exception type

* define all exception & polish compile pystack

* mapping paddle error to python exception

* polish static mode error format

* fix failed unittests

* fix dytostatic test_error

* fix check_nan_inf failed

* add unittest for coverage

* revert some code try to solve compile error

* refactor enforce & error change

* polish code & add unittest

* add sharding

* add compile limit for paddle enforce

* polish elementwise_op_function.cu.h

* fix failed unittest

* fix windows compile failed

* detail polish

* revert no type constructor

* fix strided_slice_op's GetExpectedKernelType when input tensor is at CUDAPinnedPlace

* add unittest for tensors in cuda pinned place

* skip test for cuda pinned place on cpu machines

* fix bug of fetch_async_op_handle

* revert some changes of test_buffer_shared_memory_reuse_pass

* revert some changes of test_buffer_shared_memory_reuse_pass

* nccl init support hostname and ip; test=develop