Normally I would agree, but when it comes to cuda and rocm, and the one that intel uses... you find that you have to pin torch/audio/video, and then you have to start pinning its dependencies (and their dependencies etc etc) or else you start getting random build or runtime failures...
It can be pretty bad with the bigger ML/scientific computing packages, hence all the version pinning that happens, though I will concur that generally speaking AI researchers don't understand package management in the slightest and you get all kinds of horrors as a result (e.g. mixing anaconda and pip in the same project, bonus points if it's blasted over the distro's python packaging for even more chaos. works on one very broken machine if you are lucky and don't touch anything).
Debian's approach is similarly frustrating: they will try to make your code work with an older version of a library than the one you developed with, which doesn't even work within semver (semver only defines backwards compatibility, not forwards). That's what caused the problem in this case: rolling forward dependencies was not recommended but didn't break anything, rolling a dependency backwards meant it stopped working, which is something absolutely no-one should be surprised or upset by (at least it meant the build broke as opposed to debian shipping a broken package like would have likely happened with python. It just meant that debian was shipping a version with bugs fixed in upstream). If you're going to do that you've got to be prepared to take responsibility for fixing the problem that you've introduced.
For Rust it is language specific, because the thing being statically built, doing differently is very hard.
For other languages like Python, there is not a need and you should not do that. The dynamic part of it make it so that developer should be resilient to different versions. And python also provide a lot of conveniences for that (just think about six).
But in your case, you faced the current situation of lot of "data scientist" jumping the band wagon of AI are creating a lot of things like apprentice sorcerers without a real experience of software engineering.