[llvm] [NVPTX] Customize getScalarizationOverhead (PR #128077)

[Artem Belevich via llvm-commits]

Artem-B wrote:

Sorry, I've got tunnel-visioned into looking only at the conversion to/from f16x2 vectors only. That part will probably remain the same regardless of how we do it on PTX level.

However, if the change does allow us to vectorize more scalar 16-bit ops, that is useful.

https://github.com/llvm/llvm-project/pull/128077