Test the pieusb backend in CI, and ship it on Windows

[marcinz606]

The pieusb backend landed in #747 with tests that never run and a Windows build that never contains it. Three config fixes, no behaviour change.

CI never ran the pieusb tests

ci.yml installed only the dev group, so both pieusb test files importorskip'd away and the suite reported green while covering none of the backend. pieusb is pure Python plus a bundled libusb — no system package needed.

before: 0 passed, 2 skipped
after:  20 passed, 1 skipped

tests/scanners/test_pieusb_autoexposure.py is the remaining skip and CI cannot fix it: it imports pieusb.calibration, which does not exist in pieusb==0.3.7, the pinned floor. It also tests only upstream library internals, no NegPy code. Worth deleting separately.

Windows shipped the Scan tab with no backend

The Windows release step ran uv sync --no-group scanner, which was a no-op: there is no default-groups setting, so uv installs dev alone and scanner was never in to begin with. pieusb was left out with it. Since #747 removed the _ScanUnsupportedPlaceholder gate, Windows shipped a Scan tab where both backend factories raise ScannerUnavailable.

Now uv sync --all-groups --no-group scanner — everything except SANE, which genuinely has no Windows build.

PyInstaller missed the libusb pieusb needs

libusb_package's own PyInstaller hook does collect_dynamic_libs(..., destdir='.'), putting the bundled library at the bundle root. But get_library_path() resolves it with importlib_resources.files('libusb_package') — the package directory. Verified with two probe builds:

without --collect-all:  LIBPATH: None
with    --collect-all:  LIBPATH: .../_internal/libusb_package/libusb-1.0.so

The without build still got a backend on Linux only because ctypes.util.find_library fell through to the host's system libusb. Windows has none, so get_backend() would return None and no scanner would ever enumerate.

Also confirmed PyInstaller exits 0 on --collect-all for an uninstalled package, so a local build without the group degrades exactly like gphoto2 does — no conditional needed.

Verification

• ruff check and ruff format clean
• Full suite in the CI-equivalent env: one pre-existing unrelated failure, test_stitch.py::test_stitch_real_rgb_triplet_samples (stale load_linear_preview_rgb call signature, same on main, sample-gated so CI skips it)