Rebuild example_notebooks/ginga_widget.ipynb around the AIDA API, mirroring
bqplot_widget.ipynb cell-for-cell for the shared sections (load_image,
set/get_viewport with pixel/SkyCoord/angular fov, cuts/stretch, colormap,
catalog load/style/remove including a use_skycoord catalog and the
ipywidgets slider demo, get_image/save).

Because ginga has a real WCS, the viewport and catalog cells run for real
rather than as placeholders. The interactive tail is replaced with working
ginga-native demos -- the key-bindings table, live cursor bar, click_center,
and start_marking/stop_marking/is_marking with the points retrieved via
get_catalog -- instead of bqplot's "not available" notes.

Imports now use `from astrowidgets.ginga import ImageWidget`. Verified by
executing the notebook headless with no cell errors.

Co-written with Claude Opus 4.8

Use ImageViewerLogic.__post_init__ to set up dataclass state instead of
hand-rolling the call to _set_up_catalog_image_dicts, drop the stale
ipyevents comment, and remove the redundant image_label resolution in
load_image (the helpers resolve it internally).

Co-written with Claude Opus 4.8

Pass extra keyword arguments through to ipywidgets.VBox so custom widget
traits (e.g. layout) are honored. The deprecated use_opencv kwarg is
popped first so it still warns rather than reaching VBox as an unknown
trait.

Co-written with Claude Opus 4.8

Drop the pixel_coords_offset constructor argument, the pixel_offset
property, and the associated offset arithmetic in the mouse callbacks and
_center_on. The API-approved way to shift coordinates is through the
viewport, so this special-case offset is no longer needed.

Co-written with Claude Opus 4.8

Replace the direct ginga set_pan calls in _center_on with a call to the
public set_viewport, so all re-centering goes through the AIDA viewport
bookkeeping (including the WCS handling for SkyCoord arguments). Behavior
is unchanged; the new tests are regression coverage. The now-unused
SkyCoord import is dropped.

Co-written with Claude Opus 4.8

Drop the image_width/image_height setters; the API-approved way to resize
the view is through the viewport. The getters are kept because they are
used by _viewport_window_size and set_window_size.

Co-written with Claude Opus 4.8

Use the ginga viewer's save_rgb_image_as_file renderer instead of writing
out the JPEG-encoded widget buffer. This produces a file in the format
implied by the extension (e.g. a real PNG for a .png filename) and does
not depend on a running browser. The overwrite guard is unchanged.

Co-written with Claude Opus 4.8

Replace the open review questions about cuts limits and the colormap None
guard with the confirmed answers: get_limits is exact for any interval
type, and the colormap legitimately defaults to None.

Co-written with Claude Opus 4.8

The viewport conversion already works for a non-square viewer: ginga
scales isotropically and both set_viewport and get_viewport use the window
width as the single fov reference dimension, so the round-trip is exact
regardless of aspect ratio. Document this in _viewport_window_size and add
a regression test for a non-square viewer.

Co-written with Claude Opus 4.8

Widen the try/except to cover the whole WCS-construction block (header
serialization, ginga's header parser, and set_wcs), any of which can fail
on a malformed or exotic WCS, so a bad WCS degrades gracefully to
displaying the image without one. Keep catching Exception (not
BaseException, so Ctrl-C still propagates) but log the traceback via the
ginga logger instead of printing only the message.

Co-written with Claude Opus 4.8

Ginga's ColorDist classes are parametrized, so a parametrized astropy
stretch can be reproduced instead of losing its shape parameter. Inject a
fully parametrized distribution via rgbmap.set_dist (set_color_algorithm
only ever builds ginga defaults):

- AsinhStretch/SinhStretch map exactly onto AsinhDist/SinhDist via
  factor=1/a, nonlinearity=arcsinh(1/a) (resp. sinh).
- LogStretch/PowerDistStretch configure LogDist/PowerDist with exp=a,
  matching the curve shape to within a quantization level.
- SqrtStretch/SquaredStretch/default LinearStretch use the matching
  parameterless native dist.
- Everything ginga lacks a family for -- PowerStretch (x**a, a different
  family than ginga's 'power' = PowerDistStretch), ContrastBiasStretch,
  HistEqStretch, composites -- is reproduced exactly by a new
  _AstropyStretchDist adapter that evaluates the stretch on ginga's 0..1
  ramp. This removes the warn-and-fall-back-to-linear path.

Tests assert the configured ginga ColorDist's lookup table matches the
astropy stretch on the same ramp, that the shape parameter reaches ginga,
and that PowerStretch is not mis-mapped to ginga's 'power'.

Co-written with Claude Opus 4.8