[llvm] [NVPTX] Optimize v16i8 reductions (PR #67322)

[Pierre-Andre Saulais via llvm-commits]

pasaulais wrote:

Yes, having to keep adding DAG combines to avoid the legalizer creating less efficient code is not ideal. Could we make `v4i8` a legal type? That's probably more changes to the backend overall (and having to mark a whole lot of operations as `Expand`), but might be cleaner/easier to optimize.

https://github.com/llvm/llvm-project/pull/67322