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

* feat: static check

* refactor: replace Collective & PointToPoint with NCCLEnv

* refactor: rename to RunFnInNCCLEnv

* refactor: pass std::function by value

* correct sync behavior for XPU distributed training

XPU support event mechanism similar to cuda event, so it is advisable to
use an event to sync compute/comm streams for performance. However this
mechanism is never fully tested, and inconsistent loss/ending_epochs are
reported. Therefore, this PR replaces event sync with stream waiting as
a temporary solution.

* remove compile warning

* fix device id issue for xpu eager

xpu device id is not correctly set in eager mode, thus vars are on dev0 unless
XPUDeviceGurad is called, leading to this error message for all node rank != 0:
"NotImplementedError: (Unimplemented) Place Place(xpu:0) is not supported."

* fix typo

* fix pybind error

* refactor: update pg custom

* fix: use new api in ut

* fix: typo

* revert: recover legacy apis

* fix: add GetDeviceContext

* refactor: simplify send, recv interfaces

* refactor: rm send_partial, recv_partial, all_gather_partial

* XPU support eager mode

* add unittest for XPU eager mode

* minor bugfix

* minor bugfix, test=kunlun

* correct copyright info

* 1. remove unsed vars/funcs
2. ProcessGroupBKCL inherit from ProcessGroupStream

* bugfix for fp16 in eager mode multi-card, test=kunlun

* rebase & fix a few issues

* use new processgroup interface, test=kunlun

* fix compile issue, test=kunlun

* refactor: send, recv, send_partial, recv_partial

* refactor: rm useless const ref

* [CustomDevice] GetCCLComm add custom device support

* update

* update

* update

* fix no sync bugs

* update

* update task chain

fix: update wait chain

feat: add `GetDeviceContext` for gloo

* fix oom

* fix dev

* update

* update
Co-authored-by: NLiYuRio <liyuruijx@163.com>
Co-authored-by: NForFishes <2282912238@qq.com>

* Fix bug of reduce_sum op. When input.numel() > INT32_MAX, its result
is wrong.

* support pure bfloat16

* support bf16 linear

* update PR to pass CI

* tiny fix where_grad_kernel.cu

* Support bfloat16 type for reducer and sharding.

* Fix some bug.

* Polish code.

* Polise code.

* Add bfloat16 datatype in fill_grad kernels.
Co-authored-by: Nsneaxiy <sneaxiy@126.com>

* make dense_contents_ lazy init

* update legacy dygraph

* fix legacy dygraph bug