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

[Folkert de Vries via llvm-commits]

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

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

ah that slipped through. I wanted to verify that the promotion to f64 generates the same instructions as a 64-bit fma. That actually only happens when we truly emit a 64-bit fma. 

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