Try making a portable package that pulls pytorch with CUDA (reliably) with that set up. I don't even think it's possible as you can't add another package index when using python build.
Cross-platform applications that rely on binary libraries have always been the beginning of my Python dependency misery. It usually works fine, but when it doesn't, it's miserable.
I haven't encountered any silver ammunition for that particular gun. I haven't worked on a big python project in a while, so I'm not sure if rye would do the trick.
One annoying problem that I encountered with that approach was different binaries having different version conflicts on different platforms. Not unsolvable, but definitely unsolvable without a whole lot of irritation, and a huge PITA to maintain. It definitely makes it possible though, which is a whole lot better than nothing.