port `scatter_add` to ATen (CUDA) (#38262)
Summary:
Fixes [https://github.com/pytorch/pytorch/issues/24622 ](https://github.com/pytorch/pytorch/issues/24622).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/38262
Differential Revision: D21656729
Pulled By: ngimel
fbshipit-source-id: 63dcbf8eeaf59d8295bf4e5c8bb9d28ad165d4eb
