Support various length support for SelectedRows in GLOO::AllGather (#36637)

    In cpu parallel using gloo, add various length support for SelectedRows

[cherry-pick] Support CPU Parallel in DataParallel Interface by GLOO to speed up training (#35745) (#36605)

* User specified backend (#35745)

* remove tensordot

* mv recv to head, test=develop

* del py2 code2

* fix test timeout

* fix find_unused_parameters default value

* support control flow

* supoort sync_parameters_buffers

* fix the bug of sparse embedding

* update, test=develop

* remove data parallel scale loss & apply collective_grads

* move apply in minimize

* fix failed unittests

* add gloo initializer, test=develop

* add gloo initializer, test=develop

paddle.distributed.fleet supports dynamic graph execution.

* add dygraph parallel run interface

* polish implement & unified env property name

* add print config arg

* refactor init_parallel_env function

* Compatible with multiprocessing and launch modes

* set default trainer start port

* support run in python 2

* polish python2 support code

* remove python2 support

* refine launch import

* polish dome design details

* refactor api implemention & path

* use new method _set_expected_place

* add spawn unittest framework & mnist test

* add more unittests & doc

* fix unittest failed

* polish english doc

* self review and polish details

* refactor code by reviewer's comments

* fix unittest failed

* fix parallel_env unittest

* fix several typos

* fix error introduced when fixing typos

* add unpublic note for start_processes

* polish details by xiaoguang's comment

* verify correctly when spawn nprocs=-1

* refactor spawn & init_parallel_env design

* polish doc details

* open spawn unittests

* try to fix doc compile error

* try to fix unknown doc format error

* add skip unittest when not gpu

* add retry for prefetch

* Revert "Fix/sync barrier (#25016)"

This reverts commit be6a315f.

* reopen dist UT, test=develop

* remove fl UT, test=develop

* disable distributed UT temporary，enable it soon, test=develop

use 2 cards test=develop

modify the implementation of save_persistables and save_inference_model for fleet collective mode (#20802)

* modify the implementation of  save_persistables and save_inference_model functions for fleet collective, test=develop

* add ut, test=develop

 Add detail logs on resnet unit test

* add a base class for the Communicator
* add AsyncCommunicator Impl for async distributed training