[JIT] Optimize before inlining (#35424)
Summary:
This speeds up the inlining pass of FairSeq model from 180s -> 13s.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/35424
Differential Revision: D20657271
Pulled By: eellison
fbshipit-source-id: 7a9006858c2f1b157f5a3f36ed2b3774cc186de8
