feat(show): crop kwarg for on-the-fly bounding-box rendering - #765

Add show(crop=(xmin, xmax, ymin, ymax)) to restrict a plot to a bounding
box in the rendered coordinate system, addressing #764.

- points/shapes: cheap per-element predicate subset before drawing
  (radius-aware for circles); auto-scaled color domain taken from the full
  element so colors match the uncropped plot.
- images: rasterize only the crop window via rasterize()'s target bbox, so
  the full image is never materialized (fast at Visium HD scale) and the
  zoom keeps full figure resolution. Placement is correct without a
  transform rewrite. Contrast auto-scales over the window unless vmin/vmax
  or a norm is given.
- labels: drawn in full and clipped to the box (windowed label rendering
  needs full-element palette handling; deferred).
- exact axis limits (bypass extent-union merge, honor inverted y-axis,
  ignore pad_extent); validated single-coordinate-system.

Unit tests cover the bbox helpers, points/shapes subset, full-element
color domain, image window-rasterization, and validation errors.

Follow-up to the max code review of the crop feature.

- Rename public kwarg crop -> crop_coord to match scanpy sc.pl.spatial /
  squidpy sq.pl.spatial_scatter (interop > novelty); 4-tuple unchanged.
- Norm handling is now non-mutating: capture the full-element range and
  apply it only on the matplotlib backend. Fixes the datashader leak (it
  now autoscales over the window, correct for sum/count) and the
  pre-transfunc pin (range is taken over the transfunc'd full vector).
- _bbox_mask_shapes is radius-aware per row, so a frame mixing circles
  and polygons no longer drops edge circles.
- Multiscale images pick the pyramid level for the crop WINDOW's required
  resolution, not the full image, so a Visium-HD crop no longer upsamples
  a coarse level.
- Document that labels and method='datashader' images are clip-only under
  crop (windowed labels need full-element palette handling; deferred).
- Simplify the crop axis-limit block to read crop_coord directly.

Tests: transfunc-matches-uncropped, datashader-autoscales-over-window,
multiscale-selects-finer-level, mixed circle+polygon mask.

Previously method='datashader' images ignored crop_coord (full render +
axis-clip). Now the datashader aggregation is restricted to the window:
rasterize() supplies the correct window coords/transform and only the
sliced source is materialized, so the full image is never read — this
scales to Visium HD (prototype: peak memory 258 -> 19 MB on 8k^2, and
O(window) instead of O(image)). The datashader reduction (max/sum/...) is
preserved by aggregating the sliced source in pixel space and index-
assigning onto the rasterized grid. Falls back to full render for
rotation/shear or an empty window.

New _datashader_window_image helper in utils.py; regression test asserts
the window is rasterized at figure resolution, not the full source.

Labels were the last element still full-render + axis-clip under
crop_coord. Thread crop into _render_labels so the label raster is
windowed via the same crop-aware _rasterize_if_necessary /
_multiscale_to_spatial_image paths as images (full array never
materialized; multiscale picks the level for the window).

Colour stability: windowing drops off-window instances before colours
resolve. That is stable for numeric columns, pandas-Categorical columns
(fixed levels) and explicit palettes; a plain-string column with the
default palette would reshuffle colours as whole categories drop out of
the window, so that case falls back to full render + clip. Empty-window
guard returns before the instance-overlap/colour machinery when the box
covers no labels.

Tests: placement + empty window; no-colour-reshuffle for categorical
(windowed) and string (fallback). Full non-visual suite: 390 passed.

Cleanup from a quality review of the crop feature:
- Extract _rasterize_to_bbox and use it in both _rasterize_if_necessary
  (replacing the raster_extent/do_rasterization branching) and
  _datashader_window_image, so 'rasterize a bbox to figure resolution'
  lives in one place.
- Drop _datashader_window_image's redundant has_c_dim/x_dims/y_dims
  params (recomputed from image at the top).
- Resolve the full-element crop norm range lazily on the matplotlib
  branch instead of eagerly in the crop block, so the datashader
  (large-data) path no longer pays a full-element min/max scan it never
  uses.