Maintainability quick wins: dedup, redundant copies, O(K^2) lookup

[timtreis]

Summary

A small, low-risk maintainability sweep from a refactor audit of main. No public API or behavioral change. The bugs and larger refactors from the same audit are tracked separately in #694–#700.

Changes

• Single-source the ColorLike type alias. It was defined three times (basic.py, utils.py, render_params.py), and the render_params.py copy had silently dropped list[float] — so the alias meant different things in different modules. render_params.py (the lowest module) is now canonical; basic.py/utils.py import it.
• Drop provably-dead copies:
  • _map_color_seg: map_array(seg.copy(), …) → map_array(seg, …) (×6). skimage map_array never mutates its input (verified), so the full-label-image copy was pure waste — and it runs twice per labels render when outlines are on.
  • _render_images: per-channel img.sel(c=ch).copy(deep=True).squeeze() → .squeeze(). The entry is only read (min/max) and then replaced by a fresh array (np.full/ch_norm(…)), so the deep copy was never mutated.
  • _render_points: …flatten().copy() → …flatten(). numpy/matrix .flatten() already returns a fresh array, so the trailing .copy() double-allocated.
• _extract_colors_from_table_uns: replace the per-category list.index/in scan (O(K²) for K categories) with a single {category: first_index} dict (O(K)). Bites on large category counts.
• Dedup the "blending multiple cmaps" warning into a module constant (the two copies had already drifted by a sentence).
• Remove _get_subplots — no callers in src/ (only a test exercised it) — and its test.

Verification

• multi-channel image (cmap list + single cmap) and labels (categorical + continuous) renders are byte-identical to main (RGBA buffers compared via np.array_equal).
• Non-visual test suite: 356 passed, 1 skipped (pytest -k "not test_plot").
• pre-commit: ruff check, ruff format, mypy all pass.
• Visual test_plot_* baselines run on CI; none should change (output is byte-identical).

Out of scope (flagged during the audit, deliberately not included)

• datashader import hygiene / lazy import — spatialdata already hard-deps datashader/dask/geopandas (so the supply-chain angle is moot), the lazy import touches 36 ds. sites across 3 modules, and hard-dep-vs-optional is a real design call. Better as its own PR.
• colorbar_params TypedDict — _split_colorbar_params passes unknown keys through to matplotlib, so a closed TypedDict would wrongly reject valid passthrough kwargs; needs an API-typing decision.