avoid inlining kernel lambdas on mobile (#46249)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/46249
This saves 15kb binary size on ios and increases binary size on android x86 for 30kb. It also reduces size a bit for android arm. I've talked to Martin and we should land this since Android binary size is much less important because of Voltron.
ghstack-source-id: 114177627
Test Plan: bsb
Reviewed By: ezyang
Differential Revision: D23057150
fbshipit-source-id: 43bd62901b81daf08ed96de561d711357689178f
