perf(shapes): drop the per-render table join when coloring by a table column

[timtreis]

What

_render_shapes colored by a table column joined the element to its annotating table via _join_table_for_element, whose table[joined_indices, :].copy() does an out-of-order sparse CSR row-gather copy of the whole AnnData every render. This drops that join — the original element + table are used directly.

This is the structural follow-up ("Phase 2") to #709 (the color re-join) and #711 (na_color for unannotated). After those landed, the joined element/table are no longer needed.

Why it's safe (no behavior change)

• Color is resolved per shape by _extract_color_column (perf(color): extract one aligned column instead of copying the whole table #709), which region-masks and reindexes the table itself — it never needed the table pre-joined.
• how="left" (fix: render unannotated shapes with na_color instead of dropping them #711) already keeps all shapes with na_color for unannotated, in the element's original order — so dropping the join (which equally keeps all shapes in original order) renders the same thing.
• Audit of every table/adata consumer: the matplotlib and datashader draw paths never touch the table; _decorate_axs's adata is a dead parameter; .uns["{col}_colors"] custom colors are read from sdata[table_name] inside _set_color_source_vec, not the passed object. Nothing requires the joined/region-filtered table.
• Verified pixel-identical to the join path across partial-annotation/na_color, fully-annotated, real datashader (curio, 69k circles, categorical), and real Visium (gene) — so existing visual baselines are unchanged.
• 119 non-visual shapes/utils tests pass.

Result

visium_hne render_shapes(color=gene): 451 ms → 297 ms (1.52×). The saving is the eliminated table copy and scales with table width (larger for Xenium / Visium-HD-width tables). datashader-dominated renders (e.g. curio) see little change because the copy is a small fraction of their cost.

Scope

Fill path only — the outline path keeps its own (rarer) _join_table_for_element (separate outline table + its own length-alignment logic); converting it is a small separate follow-up. _join_table_for_element itself is unchanged (still used by the outline path).