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

[Justin Fargnoli via llvm-commits]

https://github.com/justinfargnoli commented:

This seems to be a fix for an inefficient lowering of `vecreduce_fadd`.

NVPTX isn't custom lowering `vecreduce_fadd`, correct? If that is the case, then this is an issue other targets could/do face. 

Given that, why not improve the `vecreduce_fadd` lowering or implement this in DAG combine? 

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