[llvm] fix `llvm.fma.f16` double rounding issue when there is no native support (PR #171904)

[Folkert de Vries via llvm-commits]

folkertdev wrote:

Fun, fought my way through the CI failures. I'm now specifically excluding `bf16` because it's a can of worms I don't really want to deal with. There is a separate issue for it already (https://github.com/llvm/llvm-project/issues/131531).

With that, I think the fix here is good but it does not cover all architectures, notably s390x and powerpc. @uweigand do you see some way of figuring out what the systemz backend does and how to make it properly promote to `double` when it encounters a 16-bit fma?

https://github.com/llvm/llvm-project/pull/171904