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

[Artem Belevich via llvm-commits]

Artem-B wrote:

On the second thought we may be papering over the real problem that LLVM right now ends up generating rather slow code when we need to do anything with `v4i8`. If we improve that, then special-casing lowering of `v16i8`/`v8i8` etc will become unnecessary. 

I've started working on this, I should have an idea withing a day or two whether the general improvement in v4i8 lowering would be sufficient to address this particular scenario, too.

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