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

_render_shapes joined the element to its annotating table via _join_table_for_element, whose
table[indices, :].copy() does an out-of-order sparse CSR row-gather copy of the whole AnnData
(~150 ms on Visium-width tables, more on Xenium/Visium-HD). After #709 color is resolved per shape
by _extract_color_column (region-masks + reindexes the table itself) and after #711 the join is a
left join, so the joined element/table are no longer needed: use the original element + table.

An audit confirms no downstream consumer needs the joined table - the matplotlib/datashader draws
never touch it, _decorate_axs's adata is a dead parameter, and .uns custom colors are read from
sdata[table_name] inside _set_color_source_vec. Output is pixel-identical to the join path (verified
on partial-annotation/na_color, fully-annotated, real datashader (curio 69k) and real Visium), so the
visual baselines are unchanged. Measured: visium_hne render_shapes(color=gene) 451 ms -> 297 ms (1.52x).

The outline path keeps its own (rarer) join; converting it is a separate follow-up.

Author: timtreis

src/spatialdata_plot/pl/render.py

@@ -600,8 +600,8 @@ def _render_shapes(
 
     _check_obs_var_shadow(sdata, element, col_for_color, render_params.table_name)
 
-    # filter_tables=False: join_spatialelement_table below overwrites the table,
-    # so the cs-level sparse copy is wasted work.
+    # filter_tables=False: the annotating table is read directly below (color is resolved per shape
+    # from it, region-masked and reindexed), so the cs-level sparse table copy would be wasted work.
     sdata_filt = sdata.filter_by_coordinate_system(
         coordinate_system=coordinate_system,
         filter_tables=False,
@@ -610,12 +610,11 @@ def _render_shapes(
     table_name = render_params.table_name
     if table_name is None:
         table = None
-        shapes = sdata_filt[element]
     else:
+        # No join/copy: _set_color_source_vec resolves each shape's color from the table (region-masked
+        # and reindexed to the element), so unannotated shapes keep their place and render with na_color.
         _check_instance_ids_overlap(sdata_filt, table_name, element, sdata_filt[element].index)
-        joined_element, joined_table = _join_table_for_element(sdata, element, table_name)
-        sdata_filt[element] = shapes = joined_element
-        sdata_filt[table_name] = table = joined_table
+        table = sdata_filt[table_name]
 
     shapes = sdata_filt[element]