[jiterator] Move jitted_gpu_kernel into into its own header (#71960)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/71960
Changing `Loops.cuh` requires essentially every GPU kernel to be
re-compiled and so is very expensive. Since jiterator is only used by
a few files at this point, it makes much more sense to have a separate
header for it.
Test Plan: Imported from OSS
Reviewed By: mruberry
Differential Revision: D34087857
Pulled By: ngimel
fbshipit-source-id: 7caee705452a32ae34e5fcd365c0bc3ecdce73ca
(cherry picked from commit 3513aa711e09cd5a4d155b7d882e448e1a43456a)