Fix the processing logic of bernoulli on amd (#40001)
Summary:
- Fixed the bug discussed in https://github.com/pytorch/pytorch/issues/38558
- This PR is aim to make the processing of bernoulli on amd can move to the default version, even though `AT_MKL_ENABLED` is setting to `TRUE`.
- This logic used to be in the old code, but was broken by the latest update, this pr will be the fix for that.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/40001
Differential Revision: D22037646
Pulled By: pbelevich
fbshipit-source-id: c0aa4ba37416d2568daf3463cfede6838ffaeac1
