[llvm] Reland "[NVPTX] Add folding for cvt.rn.bf16x2.f32" (PR #116417)

[Artem Belevich via llvm-commits]

Artem-B wrote:

> Yea, inconsistent argument ordering in PTX does make things tricky. I've switched the patterns in this MR to use `lo` and `hi` as you suggested.

Tests could use the same treatment -- without looking up the docs for nvvm intrinsics it's hard to tell where a/b or f1/f2 are supposed to end up in the instruction arguments. While now I can be confident that the tablegen pattern is correct, I still would have hard time telling if the test patterns check what we want them to check.


https://github.com/llvm/llvm-project/pull/116417