Route GUI Qt imports through qtpy, keeping PySide6 preferred

Scope

Every Qt import in deeplabcut/gui/ goes through qtpy instead of PySide6, and the binding is negotiated at import time rather than forced.

This retains PySide6 as the preferred binding and a hard dependency which is installed with the GUI extra.
No GPL package is added to dependencies.
qtpy is MIT licensed, and does not bundle PyQt5 or PyQt6.
Users can now select an independently installed Qt binding of their choice.

Adapts the similar patch from conda-forge by @hmaarrfk, thanks again!

Mentioned in #2887, closes #3356.

• Ran GUI tests locally

Motivation

DeepLabCut hard-coded PySide6 in every GUI import and forced QT_API=pyside6 at package import, so it could not share a process with a Qt stack another package had already selected.
Routing through qtpy lets the binding be negotiated instead of dictated, which broadens compatibility without changing what DeepLabCut requires or prefers.

Main changes

• 26 GUI modules import from qtpy rather than PySide6; swap only aside from widgets.py, where Qt moves to qtpy.QtCore because PySide6 re-exporting it from QtGui is binding-specific.
• deeplabcut/gui/__init__.py selects the binding by precedence: an explicit valid QT_API, then a binding already imported in the process, then PySide6, then qtpy's own autodetection.
• The selector normalises os.environ["QT_API"] to qtpy.API after selection, so matplotlib's independent detection resolves the same binding instead of loading a second one.
• When no binding can be imported, the raised ImportError reports whether PySide6 is absent or present-but-unloadable, prints the interpreter in use, and links the installation docs. Made warnings and errors more informative and user-friendly.
• An unsupported QT_API value raises a ValueError naming the valid values, in place of qtpy's PythonQtValueError, whose message reports the variable name rather than the offending value.
• matplotlib.backends.backend_qt5agg is replaced by backend_qtagg, which exports the same names without setting matplotlib's _QT_FORCE_QT5_BINDING.
• deeplabcut/__main__.py determines GUI availability by importing deeplabcut.gui rather than PySide6, and prints that module's error rather than a fixed message.
• tests/gui/ skips on qtpy.QtWidgets rather than PySide6; the submodule is used because qtpy can be installed with no binding present, in which case the tests should skip rather than fail during collection. This is mostly to match actual qtpy imports in the code.
• qtpy>=2.4 is declared in the [gui] extra instead of arriving transitively through qdarkstyle or napari.
• Important notes regarding dependencies:
  • backend_qtagg requires matplotlib 3.5, so the floor is raised from >=3.3 to >=3.5. The resolved version is unchanged (3.8.4); the recorded specifier in uv.lock is updated to match. The <3.9 ceiling is untouched — Matplotlib v3.9 compatibility #3394 is the code preparation for lifting that end.
  • uv.lock resolves PySide6 6.9.3, due to the Linux split in pyproject, whereas 6.11+ is installed via pip install. uv does not fork and requests 6.9.3 in the lockfile.

Additional context

• deeplabcut/__main__.py is in scope because this change makes deeplabcut.gui raise where it previously could not. The entry point probed PySide6 and then imported deeplabcut.gui.launch_script unguarded. import PySide6 succeeds on the Python package even when the compiled QtCore cannot load, so that probe passes and the unguarded import reaches the new selector, producing an uncaught traceback. It also means the new installation guidance reaches users running dlc, rather than only callers importing deeplabcut.gui from a script.
• ImportError rather than ModuleNotFoundError is caught throughout, because qtpy raises QtBindingsNotFoundError, which subclasses the former but not the latter.
• Three defects predating this branch are fixed where the migration touches them: qdarkstyle.load_stylesheet_pyside6 does not exist in the pinned QDarkStyle 3.1, so launch_script.py raised on every launch and fell through to the PySide2 code path; window.darkmode() requested the PySide2 stylesheet unconditionally while lightmode() did not; and the removed lite-install message rendered as pip install'deeplabcut[gui]''.
• The qdarkstyle calls pass palette=. The zero-argument load_stylesheet() and the load_stylesheet_<binding>() helpers write os.environ["QT_API"], which would leave a value inconsistent with the loaded binding. The resulting stylesheet is byte-identical to the previous one, so the theme is unchanged.
• tests/gui/test_task_error.py patched QMessageBox on the QtWidgets module, which never reached window.py because that module binds the name directly at import. The test constructed a real modal dialog and blocked. It is fixed here because the suite cannot otherwise complete. I think Qt versions differences are why this did not surface earlier, @deruyter92 let me know if you can reproduce (before a419a77)