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

[Matt Arsenault via llvm-commits]

arsenm wrote:

> the test failures are kind of annoying
> 
> the `Generic` tests sort of require all backends, and I don't really want to build _everything_. The AMDGPU one succeeds locally, but also its name suggests it deals with `bf16`, so maybe it actually should not change at all?

Don't add tests to Generic. There's not really such thing as a "generic" codegen test; it's just a source of build pain. There's always a target dependency 

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