[llvm] [AMDGPU] Support tfe operand in image_atomic instructions (PR #92469)

[via llvm-commits]

b-sumner wrote:

I'm not opposed, but I doubt that some runtimes can properly handle TFE=1.  Do we know, e.g. that it is valid for amdgcn-amd-amdhsa?

https://github.com/llvm/llvm-project/pull/92469