Add support for reduce-scatter in c10d (#18844)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18844
ghimport-source-id: c6b2f0032c7c2212be2000a9c1f262f63d878a97
Stack from [ghstack](https://github.com/ezyang/ghstack):
* **#18844 Add support for reduce-scatter in c10d**
* #18820 Refactor ProcessGroupNCCL collective primitives
Reviewed By: mrshenli
Differential Revision: D14768369
fbshipit-source-id: a9def7a0da6e9cd995e982371cc1e22f3df1a156