* Add pure fp16 training with master weights. (#27712)

* add the weight decay func for the momentum op

* Add the multi_precision function in Momentum Optimizer.

* Make sure that the initial value of master weights are same with the fp16 weights.

* add static loss scaling.

* add the rescale_grad function in the pure fp16 training.

* use the original momentum updating method.

* Polish some codes, such as variable names.

* add docstring for apis.

* update the var creation details of _create_master_weight.

* not modify codes about imperative momentum updating.

* Fix the error of test_dist_sparse_tensor_load_momentum UT.

* add unit test for multi precision fp16 training.

* add more unit tests for CI.

* Use lower threshold values for allclose comparing in test_multi_precision_fp16_train UT.

* hot fix complle failed in gcc4.8

* fix failed unittest

* add complex64 and complex128 type; add +-*/@ and slice opreator for complex types

* add test cases for complex elementwise, matmul and getitem unittest

* add test cases for complex types

* add test cases for complex matmul unittest

* 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

- Make sure that oneDNN memory descriptors are created only once at
first iteration.

* fix paramete prefetch & device guard
Co-authored-by: NMrChengmo <cmchengmo@163.com>
Co-authored-by: Nchengmo <chengmo@baidu.com>

* 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.

* test=develop, optimize async prefetch

* fix failed tests in yingchun gived list

* add unittests into static_mode_white_list

* add enable static

* fix dist unittest

* skip test_sigmoid_focal_loss_op & add gym

* revert no need skip unittests

* remove gym

fix gru gcc7.4 bug for the gru compile

* 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

update expand as op to use the shape of the target tensor instead of the target tensor itself. (#29020)

* update, test=develop

Add eigen gru and fix the dropout bug in the rnn 

* remove -DSUPPORTS_CUDA_FP16 in cuda.cmake

* comile with cuda9

* add some unittest

* notest;test=coverage

* add unittest for trt plugin swish && split

* update ernie unittest

* fix some error message

* remove repeated judgement of CUDA version in mbEltwiseLayerNormOpConverter

* fix comile errror when CUDA_ARCH_NAME < Pascal"

* fix comile error

* update unittest timeout

* compile with cuda9

* update error msg

* fix code style

* add some comments

* add define IF_CUDA_ARCH_SUPPORT_FP16

* rename IF_CUDA_ARCH_SUPPORT_FP16 to CUDA_ARCH_FP16_SUPPORTED

Fix ops doc for some ops 

* split train_mode and has_grad

* fix format

* fix ci problems

* fix sample code

add API serialize_program, serialize_persistables, save_to_file, deserialize_program, deserialize_persistables, load_from_file. (#29034)

* Add bf16 pool2d and unify bf16 unit tests

* Add change default ops test

* Fix cpu_bfloat16_pass

* Add output_format

* Fix incorrect SetOutput

* Change fromating

* fix win ci failure, test=develop

* add ci test, test=develop

* default not show cpp statck & add hint

* fix failed unittest

* fix failed unittests

add uint8 for reshape operator

* fix tensor detach to zero copy

* fix tensor detach to zero copy

Simple code clean up

remove eigen threadpool for the speed up