Fix fake_tensor to_copy meta dispatch (#85337)
Previously, no_dispatch() was causing us to hit real kernels (well, real decomps and prims) for to_copy when we were operating on FakeTensors.
This change helps us hit meta kernels and seems to pass the relevant tests.
I still have questions about why this line has to call .to("meta")
input = new_kwargs.pop("input").to("meta")
But that can wait for another PR.
Fixes #ISSUE_NUMBER
Pull Request resolved: https://github.com/pytorch/pytorch/pull/85337
Approved by: https://github.com/eellison