[llvm] [NVPTX] add combiner rule for final packed op in reduction (PR #143943)

[Princeton Ferro via llvm-commits]

Prince781 wrote:

I ported this to DAGCombiner, but the optimization only winds up benefiting NVPTX. What do you guys think? Should we keep it in NVPTXISelLowering?

https://github.com/llvm/llvm-project/pull/143943