* add the support for multi-node training

* fix merge3 test=develop

* add paddle.fleet.AsyncOptimizer
Co-authored-by: Ndongdaxiang <dongdaxiang@baidu.com>

* support multi-ps training mode for fleetrun; test=develop

* test=develop,test=document_fix, remove the out args

* fleet_util move to paddle.fleet
Co-authored-by: NWuHaobo <wuhaobo1994@gmail.com>
Co-authored-by: Ntangwei12 <tangwei12@baidu.com>

* move dataset to fleet
test=develop

* fix
test=develop

* fix
test=develop

* fix
test=develop

* test=develop

* test=develop

* test=develop

* test=develop

* test=develop

* test=develop

* test=develop

Add dgc to fleet meta optimizer, rm dgc from optimizer all

* generate context during compile

* add fleetrun command for distributed running; test=develop

* split meta optimizer files
* add graph execution in execution, update two properties in DistributedStrategy, unit tests for these features

* add localsgd meta optimizer

* add pipeline optimizer

* add gradient Merge optimizer to meta, test=develop

Integrated Trainer of Parameter Server (API add `fluid.contrib.layers.sparse_embedding` only) (#22957)

* Integrated Trainer of Parameter Server

* refine strategy compiler and meta optimizers
make async as a_sync

* add more settings for distributed strategy
Basically, DistributedStrategy has several parts of configurations:
- BuildStrategy: the same as paddle.fluid.BuildStrategy, but the distributed arguments are moved out of BuildStrategy
- ExecutionStrategy: the same as paddle.fluid.ExecutionStrategy
- collective communication configs: nccl_comm_num, hierarchical allreduce and so on
- distributed algorithms: async_update(mainly used in PS), lars, lamb and so on

* add fleet distributed metrics
* test=develop

* fix gen nccl id bug

refactor fleet api under paddle.fleet
update DistributedStrategy

test=develop

* add paddle.fleet.DistributedStrategy for 2.0