[llvm] AMDGPU: Add round-to-odd rounding during f64 to bf16 conversion (PR #133995)

[Changpeng Fang via llvm-commits]

changpeng wrote:

> LGTM but we shouldn't really need anything custom here

Thanks. It seems we have a lot of work to do with f64->f16 (Custom) and f64->bf16 (Expand).  Just that we do not actually tahe advantage of the existing f32->f16 (or f32->bf16 for gfx950+) hardware instructions, and generate much longer sequence of code (without unsafe fast math flag set). 

https://github.com/llvm/llvm-project/pull/133995