migrate `gather` to ATen (CUDA) (#37659)
Summary:
Fixes [https://github.com/pytorch/pytorch/issues/24567](https://github.com/pytorch/pytorch/issues/24567).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/37659
Differential Revision: D21504432
Pulled By: ngimel
fbshipit-source-id: baeb464a511236b01b69a7dddd4a3db268cd799c