[fake_impls] Fix seed/offset device for attention kernels (#120839)
1) Fix fake_impls to return the correct device for these attention
kernels.
2) Remove special-casing and test file xfails
Pull Request resolved: https://github.com/pytorch/pytorch/pull/120839
Approved by: https://github.com/drisspg
