Additive RGB color blending for render_shapes / render_labels (and render_points)

Motivation

Spun out from #321 (the thread there mixes two unrelated asks — this issue is the second one, originally raised by @brainfo).

`render_images` already supports multi-channel additive compositing: you pass several channels with per-channel colormaps and they blend into a single RGB image, mirroring napari / ImageJ / FIJI for fluorescence microscopy. The same idiom is useful for sequencing-based spatial data, where a user wants to visualize co-expression of 2–3 genes per cell/spot/region on a single axis — red for gene A, green for gene B, blue for gene C, additively blended so co-expressing cells appear yellow/cyan/white.

Today this requires manual post-processing (the user in #321 mentions doing it in Illustrator).

Proposed API

Mirror the existing `render_images` multi-channel convention by letting `color` accept a list of obs/var columns on `render_shapes`, `render_labels`, and `render_points`:

```python
sdata.pl.render_shapes(
color=["Sox2", "Pax6", "Tbr2"],
palette=["red", "green", "blue"], # one color per channel
channels_as_legend=True,
).pl.show()

sdata.pl.render_labels(
color=["Sox2", "Pax6"],
palette=["red", "green"],
).pl.show()

sdata.pl.render_points(
color=["Sox2", "Pax6"],
).pl.show()
```

Output: a single axis (no multi-panel grid). Each shape / label / point gets one composited color derived from its per-column values blended via per-channel colormaps, identical in spirit to the additive multi-channel path already in `render_images`.

Why this is feasible

• The additive-blending machinery already exists in `pl/render.py` for the multi-channel image path (see `render_images` branches 2A/2B/2C). The same per-channel cmap + sum + clip logic applies.
• Single-axis output sidesteps the hierarchy-of-axes complexity that made multi-panel `color=[...]` (the other ask from #321) undesirable.
• `channels_as_legend` already exists for `render_images`; can extend.

Scope

• In scope: 2–3-column color lists on `render_shapes`, `render_labels`, `render_points`. Per-channel `cmap` or `palette`. Additive blend on a single axis.
• Out of scope (defer or decline):
  • Multi-panel grids (`sc.pl.umap`-style) — already declined in #321.
  • Arbitrary blend modes beyond additive — start with the existing `render_images` semantics.
  • PCA-based or learned color reductions for >3 channels.

Edge cases to design for

• Length mismatch between `color` list and `palette` list.
• Mixing categorical and continuous columns in the same `color` list (probably reject with a clear error).
• NaN handling per column (skip vs zero vs error — likely error, matching the recent `render_images` NaN rejection).
• Per-column `norm` / `vmin` / `vmax` — accept a list, parallel to the per-channel `norm` already supported on `render_images`.
• `groups` semantics when `color` is a list — likely ignored with a warning.

Relation to #321

#321 conflates two asks under "plotting multiple genes":

1. Multi-panel grid for several genes — declined; possibly hosted in a future Squidpy 2.0 wrapper.
2. Single-axis additive blend for co-expression visualization — this issue.

The documentation sub-tasks in #321 (`save`, `ncols` on `pl.show()`) are already resolved.

Start in pl/render.py by reading the existing render_images multi-channel branches 2A/2B/2C, then trace render_shapes, render_labels, and render_points. Done means supporting 2–3 continuous columns as per-channel colors on one axis with additive blending, while defining the listed palette-length, mixed-type, NaN, normalization, and groups behaviors.