[cherry-pick][hybrid performance] Grad fuse for gradient merge under pipeline mode (#35004) (#35299)

[cherry-pick][Hybrid Performance] Move the cast op of AMP which cast fp32 param to fp16 param to the optimizer (#34965) (#35296)
Co-authored-by: NWangXi <wangxi16@baidu.com>

* add trainer desc config to distributed strategy

* code style modified

* modify API nn.Bilinear's doc, test=develop

* fix precision of mp

* fix bug of seed

* fix dp

* print group

* fix find_unused_parameters default value

* add raw program, test=develop

* update

* sharding: update config DOC

* update pipeline config

* sharding update doc

* add model parallel support in dygraph

* support control flow

* supoort sync_parameters_buffers

* fix the bug of sparse embedding

* update, test=develop

* add tensor_indices in AssignGroupBySize

* add rebuild group in reducer

* fix fleet api doc

* add lars to fleet meta optimizer

* add lamb to proto

* add lamb to fleet meta optimizer

* fixed syntax bug

* fixed syntax bug

* fixed syntax error in lamb, add config setter of lamb in distributed_strategy

* trigger unitest to rerun

* add new unitest func for lamb

* revise unitest for lars and lamb

* revise dgc meta unitest

* revise lars document in distribute_strategy

* revise lars lamb document in distributed_strategy.py

* revise lars lamb document in distributed_strategy.py

* add weight decay exclude logic to lars

* restore optimzier.py

* restore optimizer.py as develop except lars

* add epsilon and exclude fn to distributed_sttrategy

* add lars epsilon

* revise unitest for fleet lars and lamb

* revise lars lamb unitest for CI coverage

* revise lars argument api

* revise lars argument api

* revise lars argument api

* revise api doc of lars

* fix op role

* add sharding save and add_sync_comm_for_test function

* add comm_analyse to utlis

* revise sharding_utils

* add sharding saving unittest

* revise sharding utils for unittest

* revise sharding en doc

* update sharding utils api

* add doc for sharding

* fixed bug in sharding var size count

* update varsize count in sharding

* fix sharding num_nccl_comm

* Revert "fix sharding num_nccl_comm"

This reverts commit d51587c15e9323acf226ddd36154275f0d1daf76.

* upgrade comment string to raw string

* fix string in

* fix string with ' '

* revert update on comments

* upgrade only necessary

* fix sample code checker

* fix comments with '''

* add sharding

* add get final strategy for user to print final strategy

* add adaptivelsgd

* Todo fix the code to avoid the conflict.

* rm auto from localsgd

* refine launch and distributed repr string for print

* add auto parallel L1 implementation
test=develop

* fix doc

* test=develop

* test=develop

* test=develop

* test=develop

* test=develop

* test=develop

* test=develop

* test=develop

* fix doc, test=develop

* update localsgd doc
test=develop

* update localsgd doc
test=develop

* fix fleet dgc amp doc, test=develop

* fix, test=develop

* fix async configs
Co-authored-by: Nliuyi05 <gavin1332@gmail.com>
Co-authored-by: NWangXi <wangxi16@baidu.com>
Co-authored-by: NseiriosPlus <tangwei12@baidu.com>

* fix api document

* modified timeout value on windows and mac (#26690)

* add Local SGD algorithm referenced paper
test=develop

* add cudnn related strategies to DistributedStrategy