Add CI envs for PyQt 6.5-6.7, PySide 6.5-6.7 & Python 3.12; fix Readthedocs and matplotlib 3.9, Qt6 failures

[dhomeier]

Description

WIP

[dhomeier]

Same #16 failures on ScatterRegionLayerArtist.enabled.
But there is a new one with Python 3.12 (for PyQt6?) hanging apparently in glue_qt/viewers/matplotlib/tests/test_data_viewer.py:test_add_data_with_subset
until CI timeout; I could not reproduce this latter issue locally.

[dhomeier]

@astrofrog it seems I cannot cancel workflows here (might not have the right permissions?), so would rather not experiment more in the CI at this point (haven't tested any py312+PyQT 5 combinations)

[dhomeier]

@astrofrog I think #16 may be closed now through glue-viz/glue#2479, but there are 2 new errors probably introduced with glue-viz/glue#2480, e.g. in _update_attribute_display_unit_choices:
https://github.com/glue-viz/glue-qt/actions/runs/8758215189/job/24038566338?pr=17#step:10:687
and on calling Unit(None) from a session file
https://github.com/glue-viz/glue-qt/actions/runs/8758215189/job/24038566338?pr=17#step:10:430

Also all py312 jobs still hanging in test_data_viewer.

[astrofrog]

@dhomeier - could you rebase to pick up the fixes in #21? (doesn't fix everything but at least a few jobs)

[dhomeier]

@dhomeier - could you rebase to pick up the fixes in #21? (doesn't fix everything but at least a few jobs)

I haven't followed the CI updates closely – has the default macOS-latest runner changed to arm64 since? Don't recall that many segfaults from before anyway, but the newer PyQt6 versions seem to be more stable. Same 2-4 failures though, where test_table_widget and test_table_preserve_model_after_selection are new as of PyQt 6.6.

[dhomeier]

Well ,the first common failure rather seems to be a matplotlib thing, but I am struggling to see how this could ever have worked even with mpl 3.5
https://github.com/glue-viz/glue-qt/actions/runs/9940878742/job/27458723480#step:10:283
– set_xy interface has not changed there, would this mean it has not been using a Rectangle in the older setups??

Actually the second failure assert err.strip() == "" is linked to the same
ValueError: too many values to unpack (expected 2).

[dhomeier]

I think it might indeed be axvspan changed from using Polygon to Rectangle in matplotlib/matplotlib#26788!

[dhomeier]

Darn, forgot the macos-qt65 job is hanging in

Matplotlib is building the font cache; this may take a moment.

until timeout!

I can "fix" the wrong PyQt6.QtCore.QModelIndex.row() for higher PyQT6 version by inserting an extra press_key(Qt.Key_Down), but it does not even seem to be tied to a specific version – on my Mac needing it from 6.2 onwards, but in the Windows CI it still seems to sometimes pass with 6.5, sometimes not...

[dhomeier]

@astrofrog the two remaining failures are now worked around in a rather hackish manner, that's as much as I could come up with (I tried with a wait in process_events(), but to no avail).
The macOS 14-PyQt5 test could perhaps be saved by skipping TestGlueDataDialog and TestLinkEditor tests, but that would require a bunch of system-version and architecture conditionals.
Re-enabled a test that was probably mistakenly skipped on Qt6; perhaps requires_pyqt_gt_59_or_pyside2 should be updated upstream, or maybe the Qt test helpers should actually be moved into this repo?
PySide6 is now at 6.7.2; haven't looked closely at those failures.

[astrofrog]

@dhomeier - regarding the Qt test helpers, yes they should be moved to this repo, would you be able to do that here?

I think the hackish fixes will have to do for now!

[astrofrog]

Can you rebase to take into account #12, as this might fix some of the issues?

[dhomeier]

@astrofrog adding a wait to process_events() seemed to resolve one test, but I am leaving the hacks for the remaining ones.

The QT_anything helpers could be removed from glue core at some point. Maybe make_marker should also be completely defined here, so the glue-core version would not need the extra PyQt5/6 foo?