refactor: collapse duplicated orchestration across the _render_* quartet

[timtreis]

Closes #698.

Collapses the duplicated pipeline glue across the four _render_* functions behind three seams. Pure refactor — behaviour- and pixel-preserving.

Seam 1 — datashader orchestration (highest leverage)

Extracted _shade_datashader_aggregate() into _datashader.py: the core every datashader caller repeats — _ds_aggregate → _apply_ds_norm → transparent-NA guard → _build_color_key → categorical/continuous shade dispatch. _render_shapes (previously inline) and _datashader_points now both call it, keeping only their element-specific prep (geometry transform vs point parse + spread_px) and their own outline / image / colorbar steps.

This is the seam with real maintenance history: #367/#687 (double-alpha), #372/#376 (clip/cmap-alpha), #617 (_apply_user_alpha) each had to be applied to both copies. Now there is one.

Note: since the issue was filed, points was already extracted into _datashader_points (via #714), so this seam unifies shapes-inline ↔ _datashader_points rather than two inline copies.

Seam 2 — shared preamble

Extracted _warn_groups() and _maybe_apply_transfunc(), now used by shapes, points and labels. The filter_by_coordinate_system / groups-filter blocks the audit mentioned are not actually identical across the three renderers (GeoDataFrame vs points+adata vs label-id masking), so those are intentionally left in place rather than forced into a false abstraction.

Seam 3 — labels legend path

_render_labels re-inlined _add_legend_and_colorbar's body and had drifted on na_in_legend and the categorical colorbar term. Generalized the helper with is_continuous_override / na_in_legend_override and routed labels through it, deleting the ~50-line copy. The overrides reproduce labels' exact current behaviour.

Verification

• Non-visual suite: 415 passed, 1 skipped.
• Visual test_plot_* (shapes/points/labels) fail identically with and without this change locally — same 217/131 split, byte-identical failing sets — confirming the diff doesn't move any rendered output. (Local visual failures are pre-existing matplotlib-vs-CI-baseline drift, unrelated to this PR.)
• ruff check / ruff format clean.

No public API, default behaviour, or I/O changes. No new dependencies.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 77.13%. Comparing base (4ba13a3) to head (2761d84).

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #717   +/-   ##
=======================================
  Coverage   77.13%   77.13%           
=======================================
  Files          14       14           
  Lines        4457     4440   -17     
  Branches     1023     1014    -9     
=======================================
- Hits         3438     3425   -13     
+ Misses        661      658    -3     
+ Partials      358      357    -1     

Files with missing lines	Coverage Δ
src/spatialdata_plot/pl/_datashader.py	91.28% <100.00%> (+0.77%)	⬆️
src/spatialdata_plot/pl/render.py	88.37% <100.00%> (+0.21%)	⬆️

... and 1 file with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.