[llvm] [NVPTX] Use PRMT more widely, and improve folding around this instruction (PR #148261)

[Artem Belevich via llvm-commits]

Artem-B wrote:

Reproducer for assertion-enabled llc: https://gist.github.com/Artem-B/0fd96044926baa90ea83180f9a00c1b4
Note that it no longer crashes the llc at HEAD. It's not clear yet if the issue is fixed, or became harder to reproduce.

https://github.com/llvm/llvm-project/pull/148261