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

[Shilei Tian via llvm-commits]

shiltian wrote:

That is the problem with the generic target, but apparently we can't do anything about it. We already have many "bad" generic targets where the generic target can have a feature that doesn't exist on all of the covered targets, and the feature also can't be generalized using some merging rules like the ones proposed here. The one I'm referring to here is XNACK. To support that, we're essentially making the validator accept a feature on a generic target that is absent from a target it can cover, which is exactly where this entire conversation started.

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