[export/nonstrict] always create FakeTensorMode (#119446)
Previously in non-strict mode we would source a FakeTensorMode from existing tensors if available.
It turns out this is problematic, as it means we can't directly control the behavior of this FakeTensorMode. For example, if the user-provided FakeTensorMode does not set `allow_non_fake_inputs=True`, then we get into trouble with constant tensors, etc.
At the moment, we still have to epxlicitly re-fakifky the module state. @ezyang has recommended against this, but it's necessary because `create_aot_dispatcher_function` calls `detect_fake_mode` on all the inputs, which will error if not all the FakeTensors are on the same mode. We should straighten this out, but leaving for the future.
Differential Revision: [D53559043](https://our.internmc.facebook.com/intern/diff/D53559043/)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/119446
Approved by: https://github.com/ezyang, https://github.com/zhxchen17
