[llvm] AMDGPU: Add pattern for copysign of 0 (PR #172699)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> Wouldn't it be better to expand fcopysign earlier so that normal combines and selection patterns can do their work?

That breaks value tracking and combines that want to check if it's legal. I remember going back and forth on this and this worked out better 

https://github.com/llvm/llvm-project/pull/172699