perf: legend/decoration path re-expands the compact categorical colour vector (all renderers) #732

Following #731 (points kept colour as codes+palette): the shared legend/decoration path still expands the compact pd.Categorical (codes + tiny hex palette) to an O(n_elements) object array once per render — for points, shapes and labels.

Three spots, each only needs the k categories:

• _want_decorations (render.py:133): np.asarray(color_vector) + O(n) == scan → use .categories/.codes.
• ColorSpec.make_palette (_color.py:754): dict.fromkeys(full_vector) + pd.Categorical rebuild + n-mask → .categories.
• _decorate_axs (utils.py:493): builds an n-row DataFrame then drop_duplicates to recover a k-sized cat→colour dict that already exists upstream (_color.py:678).

Fix: read .categories/.codes (k entries) instead of materializing the per-element vector. Lowest-effort, benefits all three renderers at once. Must stay output-identical.