[llvm] AMDGPU: Make v2f64 -> v2f16 conversion Legal only when unsafe fast math is set (PR #134738)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> I understand your proposal to avoid use of global options. But here we need to be consistent. Options.UnsafeFPMath has been used in all other places to expand/lower f64->f16. We should not use a different approach only here.

This is because traditionally the FP casts did not have fast math flags. They do now, and we should migrate to only using them.

https://github.com/llvm/llvm-project/pull/134738