* add trt_support_nhwc_pass

* Fix bug of reduce_sum op. When input.numel() > INT32_MAX, its result
is wrong.

* Remove climits.

* Fix bug of paddle.save. It may cause bug for saving sharded optimizer
state_dict() in parallel.

* [Auto Parallel] Rename methods of ProcessMesh

* [Auto Parallel] Impl the python process_mesh by the c++ one

* [Auto Parallel] Add some minor modifications

* [Auto Parallel] Rename some methods

* [Auto Parallel] Remove unnecessary codes

* [Auto Parallel] Add back some removed files

* [Auto Parallel] Fix bugs

* [Auto Parallel] Fix a bug

* Update process_mesh.cc

* [Auto Parallel] Merge dist attrs of Python into C++

* [Auto Parallel] Add back deleted importing

* [Auto Parallel] Add back removed unittest

* [Auto Parallel] Remove type qualifiers of return types

* [Auto Parallel] Fix some bugs

* [Auto Parallel] Fix a bug of the quant pass

* [Auto Parallel] Fix the code style

* [Auto Parallel] Clear some fluid APIs

* support elementwise base func

* fix compiling error and add test

* remove additional param

* support vjp for div using comp

* remove additional change

* fix dy2st error with magic num

* fix dy magic num

* another magic

* another magic

* add more test

* fix windows problem

* another magic

* fix windows compile

* invoke ci

* add skip rename strategy

* support add vjp

* fix test_tanh

* support add with new axis cal

* fix resnet and some test

* add composite log

* support sub vjp

* enhance_tests

* support more dtype for full

* add fmha_flashattention oss plugin

* add where, atan2, median 0d ut

* add where, atan2, median 0d ut

* update

* update

* update

* update trt engine to set in/out data type

* update

* Update engine.cc

* Update engine.cc

* update

* set engine output type before freeze the network

* update

* update trt autoscan ut

* update

* update ut

* fix equal bug, update ut

* fix cast and equal ut

* update cast ut using TRT < 8.4

* set datatype from scope

* check output var is nullptr

* Update op_converter.h

* update tensorrt_engine_op_test ut

* update

* clear ProcessGroupCustom manually

* fix bug

* fix bug

* move destroy ProcessGroup to ProcessGroupIdMap

* enable destroy to all device

* remove unused comments

* change to internal api

* Update process_group.cc

* Update process_group.cc

* update get_device to custom

* add custom_device api

* rm is_compiled_with_custom_device from framework

* add todo comments

* support elementwise base func

* fix compiling error and add test

* remove additional param

* support vjp for div using comp

* remove additional change

* fix dy2st error with magic num

* fix dy magic num

* another magic

* another magic

* add more test

* fix windows problem

* another magic

* fix windows compile

* invoke ci

* add skip rename strategy

* support add vjp

* fix test_tanh

* support add with new axis cal

* fix resnet and some test

* add composite log

* support sub vjp

* [Eager] polish some apis logic

* polish api logic

* kunlun add support for c_concat and c_split

* replace mutable_data() and ShareDataWith()

* fix xpu unittest issue: zero_dim_tensor

* deal with leftout issue introduced by #49470

* [PHI] rrelu add yaml

* polish

* polish

* [Eager] polish some api logic

* fix split

* revover

* lerp support 0 Tensor

* fix lerp grad

* fix lerp zero test

* fix 0D + ND/ND + 0D

* fix check

* update code

* fix lerp infer shape

* static backward test

* updata static graph test

* Fix reduce func bug in process_group_bkcl

Also catch up with a recent process_group PR that failed to add XPU branch.
Note that reduce is still accomplished by allreduce for xpu. Fix this should
xccl lib be updated.

* fix compile issue for non-XPU

* compile fix

* fix compile

* compile fix

* add more preln

* [Zero-Dim] support input 0D Tensor for maximum,minimum,allclose,sigmoid_focal_loss

* [Zero-Dim] add backward test for sigmoid_focal_loss with 0-D input Tensor

* [Zero-Dim] support input 0D Tensor for fmax,fmin,complex