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

[Folkert de Vries via llvm-commits]

================

----------------
folkertdev wrote:

huh, doesn't picking the `bf16` fma libcall do the right thing? or are you saying that currently it would pick the `f32` libcall (with associated promotion/rounding)?

In any case even promoting to `f64` will not give accurate results for `bf16`, so I'm not sure what to do?

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