Change FakeTensor serialization to consider only an _active_ FakeTensor mode (#120848)
Summary: https://github.com/pytorch/pytorch/pull/108186 make some changes related to FakeTensor serialization such that saving and loading a tensor will give us a meta tensor, even if FakeTensor mode is not enabled. This means we can't properly save and load Tensors as part of Fx graph caching. This PR changes the logic to check if there's an _active_ FakeTensor mode.
Test Plan:
* New unit tests
* Validated unit tests introduced in https://github.com/pytorch/pytorch/pull/108186 still pass
Pull Request resolved: https://github.com/pytorch/pytorch/pull/120848
Approved by: https://github.com/eellison, https://github.com/thiagocrepaldi
