Fix #489: use public supervision API for annotation (works on 0.18-0.29)

[prawnsgupta]

Fixes #489.

Problem. tabs/dataset_viewer.py imported BoxAnnotator from the internal path supervision.detection.annotate, which was removed in supervision ≥ 0.22, causing ModuleNotFoundError for anyone whose environment resolves a supervision newer than the pinned 0.18.0. While investigating I also found that the shared utils/image.py::draw_detections() helper (used by tabs/inference.py) crashes on modern supervision because its palette fallback calls sv.Color.DEFAULT / sv.ColorPalette.default(), both of which raise AttributeError on recent versions (and are not caught by its except TypeError).

Fix.

• utils/image.py: use sv.ColorPalette.DEFAULT (available on old and new supervision) and keep the existing old-API/new-API try/except, so both BoxAnnotator-combined (≤0.21) and BoxAnnotator + LabelAnnotator (≥0.22) paths work.
• tabs/dataset_viewer.py: remove the broken inline annotation and delegate to draw_detections(), matching how tabs/inference.py already renders detections.

Verification. Tested end-to-end against supervision 0.18.0 (current pin) and 0.29.1 (latest); both render boxes/labels and return a valid uint8 ndarray. No dependency changes required — confirmed numpy 1.26.4 (project pin) coexists with modern supervision.

[dpascualhe]

Hi @prawnsgupta , I have just tested your PR and it works nicely! However, the GUI has been refactored and therefore your changes conflict. Could you resolve those conflicts? Basically, your changes from tabs/dataset_viewer.py should be moved to tabs/tasks/image_detection/dataset_viewer.py. Upon fixing that, we can merge.

[prawnsgupta]

Thanks for testing it, @dpascualhe, and sorry for the slow turnaround.

The conflicts are resolved. The branch is merged with current master and the
annotation changes now live in gui/tasks/image_detection/dataset_viewer.py,
following the refactor that moved tabs/ to gui/ (daa19f0).

The change is down to two files:

• gui/tasks/image_detection/dataset_viewer.py — drops the direct
  supervision.draw.color / supervision.detection.* imports and calls the
  shared draw_detections() helper instead.
• perceptionmetrics/utils/image.py — uses sv.ColorPalette.DEFAULT, which is
  valid in both the old and new supervision APIs, and passes the palette and
  text styling through to the newer BoxAnnotator + LabelAnnotator path
  instead of constructing bare annotators.

Verified locally on Python 3.10 against the locked supervision 0.18.0:
tests/gui and tests/utils/test_image.py pass (21 passed), and the full suite
is unchanged at 77 passed / 1 skipped — the 4 failures I see are the existing
Cityscapes, nuImages and SemanticKITTI dataset-fixture tests, unrelated to this
branch.

I think it's ready to be merged. Let me know for any other changes.