* gather with doc

* resolve comment

* polish

* polish

* code style

* polish doc

* add_test

* polish

* polish

* add test check

* add test check

* polish

* polish

* polish

* polish

* fix_time_out

* polish

* fix timeout

* fix_timeout

* polish

* polish

* polish

* polish

* polish

* remove distribute

* fix py3 bugs

* fix gpu-ps bugs

* fix compile bugs

* fix unittest bugs

* add gloo  send_recv

* fix code_stype

* fix CI bug

* fix send_recv.cc

* add send_recv without sync_op

* fix send_recv test

* fix gather.cc

* [kunlun] support min/max in dygraph mode

* update xccl to 1.0.13

* add fused_yaml fused_backward

* fix eager_funciton bug

* add some comment of fused yaml file

* add 'support_dygraph_mode' configuration in fused yaml

* delete some 'fused_api.h' in include file

* add fused flag in api_gen

* move DataType from paddle::experimental to phi

* convert namespace

* convert namespace

* convert namespace

* clarify namespace

* convert more datatype

* Revert "convert more datatype"

This reverts commit 083b462959e6a22d4d8767707b628b95b396642e.

* convert more in auto_code_generator

* fix conflicts for XPU

* fix namespace conflicts

* fix errors

* Revert "fix errors"

This reverts commit f9d9958b54ee32141112274c8a5c3c381ab0f876.

* fix errors

* fix formatting

* remove contexts in tensor_utils

* update from_blob

* update from_blob

* update from_blob

* fix bug

* fix bug

* [CustomDevice] fix SyncDefaultStream for process_group_custom

* update

* add UpdateWaitChain for  process_group_custom

* add UpdateWaitChain for  process_group_custom

* [CustomDevice] fix process_group_custom api

* update

* update

* update

* update

* [kunlun] support reduce_scatter

* uncomment unittest

* update xccl to 1.0.10

* update xccl lib & use native Reduce in dygraph

* minor

* revert to use default XPU stream for computing

XPUContext now has a null stream by default. If you want to use a separate stream
 (e.g. in async collective communication), you should create a dedicated XPUContext
and invoke its XPUContext::CreateStream()

* minor

1 update xccl lib
2 when using comm_ctx, the allocator should be set manually.

* clear ProcessGroupCustom manually

* fix bug

* fix bug

* move destroy ProcessGroup to ProcessGroupIdMap

* enable destroy to all device

* remove unused comments

* change to internal api

* Update process_group.cc

* Update process_group.cc

* refactor: migrate comm checks

* refactor: add check in comm context

* feat: add gloo static check

* refactor: add place param in static check

* Fix reduce func bug in process_group_bkcl

Also catch up with a recent process_group PR that failed to add XPU branch.
Note that reduce is still accomplished by allreduce for xpu. Fix this should
xccl lib be updated.

* fix compile issue for non-XPU

* comm_context and static init

* refactor: move to phi/core/distributed

* refactor: avoid mutable_data usage

* fix: windows sock

* fix: device without nccl
Co-authored-by: Wen Sun <syl1887415157@126.com>

* fix: fix hidden virtual funcs

* fix: add default impl

* refactor: use base class

* fix: incorrect deps

* fix: add missing header

* refactor: update class structures

* fix: bkcl typo

* fix: remove redundant def

* nullptr bugfix for XPU pg mode

Also a few kernels is added to xpu whitelist

* increase error msg length

* chore: unify `SingleTensor`

* feat: dynamic check

* refactor: classify static check

* refactor: rename to static_check & use forward decl

* refactor: switch to unary & binary funcs

Note: this is a temporary solution, should be replaced once reduce kernel
is natively supported on KL2

* feat: static check