[llvm] [AMDGPU] Generalize generic-target numeric property validation (PR #223177)

[Matt Arsenault via llvm-commits]

arsenm wrote:

The validation is far weaker than I wanted. The problem is anti-features, like the hazards. The obvious thing requires inverting all of these to marking most targets as not-has-hazard which is a hassle 

https://github.com/llvm/llvm-project/pull/223177