[clang] [Clang] Add __has_target_builtin macro (PR #126324)

[Joseph Huber via cfe-commits]

jhuber6 wrote:

> Marking as requesting changes so we don't accidentally land, because I've got questions about whether we want to go down this route.
> 
>     1. GCC has `__has_builtin`, so how do they handle offloading targets? Do they have the same odd behavior where `__has_builtin` returns true for builtins it cannot actually emit code for?
> 
>     2. Given that `__has_target_builtin` seems to have the semantics everyone would expect from `__has_builtin`, do we want to consider deprecating `__has_builtin` so that downstreams have time to adjust but we eventually end up with a less confusing builtin?
> 
> 
> It really seems to me that `__has_builtin` has a broken design because of compilations where we try to hide a two-pass compilation as though it were one-pass and it seems like we're going to confuse folks with that behavior. If I could wave a magic wand, I would say `__has_builtin` should behave exactly how `__has_target_builtin` is behaving in this PR and that a reliance on `__has_builtin` behaving how it does today is relying on a bug.

That's what the initial proposal / patch was, but that got reverted because it broke some CUDA on Arm sources.

https://github.com/llvm/llvm-project/pull/126324