[llvm] [CodeGen] Pass the atomic instruction to supportsUnalignedAtomics. NFC (PR #216621)

[Harrison Hao via llvm-commits]

harrisonGPU wrote:

> I'd prefer to not spread lowering hooks that depend on exact IR references, rather than properties. Particularly for the elementwise atomic case, it seems wrong to overload the meaning

Thanks, Matt. I've updated it to only take Alignment and SizeInBytes as arguments.

https://github.com/llvm/llvm-project/pull/216621