[llvm] Handle VECREDUCE intrinsics in NVPTX backend (PR #136253)

[Princeton Ferro via llvm-commits]

Prince781 wrote:

After some thought, it seems all three patterns of lowering a reduction have their tradeoffs and each is preferred in some circumstance. I'd like to propose keeping the default lowering (shuffle reduction), but add a way for clients to override this with metadata. I'm thinking about something like `!reduce.tree` on the reduction call. Before I proceed, I'd like to know if this is a plausible idea and if I should propose it in an RFC. Thanks!

https://github.com/llvm/llvm-project/pull/136253