internal fixes for FunctionalTensorMode usage in AOTAutograd (#538)

Summary:
Pull Request resolved: https://github.com/pytorch/executorch/pull/538

Fixes needed to properly land https://github.com/pytorch/pytorch/pull/110079 internally

(1) executorch has a higher order op that requires a functionalization rule

(2) s-curve export still has an internal flow that calls some AOTAutograd API's, but also manually makes some calls to C++ funcitonalization. I changed them to use python functionalization.

Reviewed By: zou3519, tugsbayasgalan

Differential Revision: D49657241

fbshipit-source-id: 7b9cc46851540200178dc4748baf5dd623f71082