Lift non-FakeTensor restriction for compile (#107042)
Currently, we have the assertion that dynamo won't accept FakeTensor input unless we're exporting. This PR try to remove this restriction to finish https://github.com/pytorch/pytorch/pull/105679.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/107042
Approved by: https://github.com/ezyang, https://github.com/zou3519