cherry-pick of #32972

cherry-pick #32561

* skip op has no fp16 kernel

* add ut

* custom python backward

* polish up the code

* polish up the code

* polish up the code.

* Fix code format and comments.

* Delete redundant files.

* add unnittest.

* edit unnittest.

* edit unnittest.

* Remove redundant header files.

* Improve coverage and remove redundant code.

* support saving for backward.

* polish code according to comments.

* Add support type for PyLayer.

* Modify the DOC.

* polish Doc.

* polish Doc.

* polish Doc.

* polish Doc.

* polish Doc.

* polish Doc.

* polish code and make the code robust.

* Modify the code format.

* add register backward hook method

* add leaf grad accumullated test

* add layer.to api

* add layer.to api

* add layer.to api

* add the doc for Layer.to

* add input type checking

* modify assert and import bug

* format code style

* format code style

* make place support str type

* add SetGradVarBase method to set the gradient after conversion

* modify argument palce to device

* modify argument palce to device

* modify doc of layers.to API

* add xpuplace to device argument

* add custom init grad for backward function

* add custom init grad for backward function

* handle when the grad_tensor is none

* handle when the grad_tensor is none

* fix the args type error on windows platform

* modify the args order and doc

* format code

* add grad_tensor to xpu

* modify the grad_tensor type check

* add paddle.backward api to support multi tensors gradient compute

* add paddle.backward api to support multi tensors gradient compute

* add paddle.atuograd module and backward api

* change tensor.backward func args

* modify tensor backward api

* remove create_graph intputs args

* add doc and examplex code for backward api

* when have the same tensor, throw error

* modify test Init func args

* modify the execute.Init func args in test files

* add paddle.autograd package in setup.py.in

* modify error msg, remove _run_backward method in class Tensor

* add test cases for backward api

* new group

* ci compatible fix

* assert nccl

* refactor and simplify hook design

* fix reducer add hook error

* add Tensor.register_hook basic impl

* refine prepare data impl

* revert prepare data change

* support register_hook for Tensor

* add hook test in model

* polish tests and doc example

* fix double grad test failed

* remove reduce hook func

* fix set empty error

* polish code by comments

* change reduce_hook to mutable_hook

* remove useless tmp_ins

* fix shape code format error

* fix shape code format error

* polish tensor pipeline. test=develop

* [static setitem] support the index step > 1. tensor_a[::3] = value

* [static setitem] support the index step < 0. Eg: tensor_a[::-3] = value

* [static setitem] support the index is Tensor. eg: tensor_a[tensor_3:0:-1] = value

* Add op version.

Performance optimization for dynamic setitem: Call op set_value to speed up because the original call to TensorToPyArray will introduce unnecessary data copy. (#30817)

* support layer_norm fp16 in dygraph amp

* add ut

* refine code

Set expected place in child thread for dataloader to avoid costing cuda memory on other card (#30338)

* set expected place in child thread for dataloader

* set device id when set tensor from numpy

* revert tensor_py change

* add compile guard

* fix ci

* fix bug

* change to tensor copy sync

* change to tensor copy sync

* make copy_to safe when use TensorCopy

* refine code

* add ut

* add cudapinned garbagecollector

* add testcase: cpu place -> cuda pinned place

1. when slice_item is a slice: 
 1) the start of __getitem__ should be std::max(start, 0) if slice
 2) the start of __getitem__ should be std::min(end, dim) 
2. when slice_item is an integer, it should be in [-dim_len, dim_len) 
3. Fix error message to use accurate data

* Revert "[inplace] Add ShareHolderWith for class Variable and SharePlaceholderWith in VarBase.detach() to share the same Tensor/SelectedRows (#29267)"

This reverts commit b10ecd9d.

* Support ShareInplaceVersionCounterWith to share the same inplace version counter for VarBase

* fix fleet for multi-stream

* fix memcpy for ncclid

* use sync to solve move operation

* support deepcopy for Layer/Tensor/Paramerbase

* fix some code

* add tensor_indices in AssignGroupBySize

* add rebuild group in reducer

[inplace] Add ShareHolderWith for class Variable and SharePlaceholderWith in VarBase.detach() to share the same Tensor/SelectedRows (#29267)

* The leaf tensor concept is exposed and the gradient accumulation of leaf tensor

* The leaf tensor concept is exposed and the gradient accumulation of leaf tensor

* fix coverage

* fix api doc

* fix CI unittest

* fix CI unittest

* fix unitest

* empty tensor does’t need inner_var_

* fix some error message

* Add a class TensorInplaceVersion to count the inplace version and put it in framework::Tensor instead of Allocation or Variable.

* Add a new attribute `_inplace_version` for VarBase.

* Raise exception if an inplace operation can result in incorrect gradient computation.

* Add a new interface _bump_inplace_version() for VarBase to bump the version whenever the Tensor is modified through an inplace operation.

* For api assign, call _bump_inplace_version() when it's an inplace operation inn dynamic mode.

* Use original var_wrapper if the inplace_version is not changed.

* Replace SnapshotVarWrapperList with SnapshotVarWrapper to optimize performane.

* add reducer

* refine envent for memorycopy

* add concat&split for allreduce

* apply concat & split for fuse tensor

* fix nccl dep

* fix the untest, compile problem and ddp initialize problem

* fix untest for mac & add some comments & solve the repeated param in sublayers

* fix untest for windows & fix document

* split train_mode and has_grad

* fix format

* fix ci problems

* fix sample code

* fix tensor detach to zero copy

* fix tensor detach to zero copy

* make Numpy version is below 1.19.3

* fix 2.0 doc

* change paddle.fluid.layers.reduce_sum to paddle.sum in sample codes

* format codes

* add tensor clone

* fix unittest test_var_base