Use reentrant mode when invoking prims, delete global prim_fake_mode (#84090)
Maybe I should be using the meta_impl instead of the prim_impl, but it's not terribly clear why, since the prim impl will be better tested and should work under the re-entrant FakeTensorMode.
Fixes https://github.com/pytorch/pytorch/issues/78613 in the process
Pull Request resolved: https://github.com/pytorch/pytorch/pull/84090
Approved by: https://github.com/ezyang, https://github.com/samdow
