[llvm] [mlgo] inline for size: add bypass mechanism for perserving performance (PR #95616)

[Mircea Trofin via llvm-commits]

mtrofin wrote:

You mean we'd mark the cold callers as minsize and drive the bypass decision that way? Yes. In fact we want to experiment with that next, and we may then remove the bypass flag if we can do it all with one flag. I was trying to be a bit conservative because minsize would also affect how those callsites are compiled post-inliner. Which should compound more for size, and matter not for speed, but, again, wanted to be conservative.

https://github.com/llvm/llvm-project/pull/95616