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

[Folkert de Vries via llvm-commits]

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

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

in this case the input is already promoted, so we get the somewhat unfortunate double promotion/demotion, e.g. in the case of wasm:

```
    call	__extendhfsf2
	f64.promote_f32
	f64.add 
	f32.demote_f64
	call	__truncsfhf2
```

is there a simple fix for that, or should we just leave it?

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