How to drop 32bit support from the scientific Python stack

[Miro Hrončok]

Hello folks,

Couple months ago, we discussed with @psimovec if it's possible to ExcludeArch 
i686 from scipy.

This Python discussion brought the topic back:
https://discuss.python.org/t/dropping-32-bit-packages/5476/9

I've tried to see how many packages would be affected and the short answer is: 
all of them, unless we break the dependency chains somewhere.

There are many deep transitive build dependency chains on scipy, but the 
perhaps most interesting one is:

(everything) <- rpm <- rust-rpm-sequoia <- rust-packaging <- pytest <- 
python-hypothesis <- python-pandas <- scipy

Another interesting fact is that the majority of the dependency chains I found 
traverses through a path of noarch Python packages (such as pytest).

If scipy drops i686 we would need to conditionalize the test-related 
BuildRequires of such noarch packages -- unfortunately that would mean that 
based on randomness (noarch packages are built on random archotcture), not all 
tests would always run and we could potentially ship packages that are broken / 
untested properly.

And if we don't conditionalize the dependencies, the packages would randomly 
fail to build.

An alternative is to make all the related packages archful with noarch 
python3-xxx subpackages -- which would be quite tedious and wasteful resource-wise.

The best way forward for this use case (and many others that will show up 
sooner or later) would be to stop building noarch packages on i686. That way, 
only archful packages that (Build)Require scipy would need to be changed.

Would that be possible?