gh-122931 CI for Linux multiarch co-installability

[stefanor]

As suggested in this thread, add some CI tooling to ensure that Python continues to be multi-arch co-installable.

Explicitly skip checking abi3.so and abi3t.so extensions, as they are currently not co-installable (#122931).

We extend reusable-ubuntu to perform an extra set of build+install+hash runs. This required some slightly different configure options to get more co-install-friendly builds. Those configure options broke some existing tests, but the fixes are trivial, and included.

From an earlier implementation that didn't use hash-manifests:

Example failure (without abi3 and abi3t ignored): https://github.com/stefanor/cpython/actions/runs/28550736484/job/84647793000
Example success (with them ignored): https://github.com/stefanor/cpython/actions/runs/28551134448/job/84648971317

• Issue: Allow stable ABI extensions to include multiarch tuples in the filename #122931

[stefanor]

Note: This will require an extra ~165MiB * 5 test builds = ~820 MiB of artifact storage in GitHub CI, per test run. These are stored for 90 days by default. We could configure less...

Duh. Let me store hash manifests instead of storing the installs.

[webknjaz]

We do not do this in the main test suite execution, as achieving multi-arch co-installability requires some non-default configure parameters.

Do you mean the debug option?

But we could share builds, if desired for performance.

Seems like having another matrix will add 15–16 minutes of CI runtime.

[stefanor]

Seems like having another matrix will add 15–16 minutes of CI runtime.

That's a lot. OK, I'll attempt to include it.

[stefanor]

We do not do this in the main test suite execution, as achieving multi-arch co-installability requires some non-default configure parameters.

Do you mean the debug option?

I meant a multiarch libdir and --with-build-details-suffix. And in fact I needed two extra commits to fix tests with these options:

• d4c6100 that I probably should have included in build-details.json should be installed to different locations for nondebug/debug builds and have distinct base_interpreter values #131372
• d390e7a which is already a PR I've already filed a while ago: Strip absolute --libdir paths from configure args in test_freeze #124916

[stefanor]

Seems like having another matrix will add 15–16 minutes of CI runtime.

It's part of the main test matrix, now.