[llvm] [AArch64][GlobalISel] Take abs scalar codegen closer to SDAG (PR #84886)

[Amara Emerson via llvm-commits]

aemerson wrote:

I agree with Matt, `lower()` should be checking for the legality of SMAX and use that if available, falling back to the current XOR method, see `TargetLowering::expandABS()`. But for AArch64 that's only legal with the CSSC extension, so you'll need to check why we generate the compare lowering in DAG at all.

https://github.com/llvm/llvm-project/pull/84886