Make ELU great again (#33244)
Summary:
Due to compiler bug, we have to make some workaround on ELU for CUDA. A necessary condition for this bug to happen is `invoke_with_array` in `Loops.cuh`. Now, https://github.com/pytorch/pytorch/issues/33222 will kill that function, and we need to remove that workaround once https://github.com/pytorch/pytorch/issues/33222 is landed.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/33244
Differential Revision: D20076197
Pulled By: ngimel
fbshipit-source-id: 39f99783014c78cecad1c39cb46092278ff220b9