Enable BFloat16 for `logaddexp` & `logaddexp2` on CUDA (#57908)
Summary:
Enabled BFloat16 for `logaddexp` & `logaddexp2` on CUDA, with a [workaround](https://github.com/pytorch/pytorch/pull/57908#issuecomment-837320532) suggested by zasdfgbnm.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/57908
Reviewed By: mruberry
Differential Revision: D28344976
Pulled By: ngimel
fbshipit-source-id: edef654b5819b236fbd9996f962115beb6e147e1