[llvm] [AMDGPU] Enhance verification of amdgcn.cs.chain intrinsic (PR #128162)

[Matt Arsenault via llvm-commits]

arsenm wrote:

This intrinsic should really be required to use callbr. There isn't a sensible way for generic IR transforms to respect this condition, it just happens it's unlikely something would want to place instructions at the end of a block ends in unreachable

https://github.com/llvm/llvm-project/pull/128162