fix(graphics): evict GPU textures when their pixels are freed

[cantona]

fix(graphics): evict GPU textures when their pixels are freed

Problem (fixes #1591)

The renderer keeps one GPU texture per standalone image in
Renderer::image_textures. That map is only ever pruned two ways:

1. the byte-budget LRU in render_images (evicts the least-recently-drawn
   textures once the cache exceeds its budget), and
2. an explicit kitty eviction (evict_image_texture).

Neither reacts to an image being freed. When a graphic goes away — an
atlas graphic's last referencing cell scrolls off the grid, or a kitty
image is deleted — its CPU-side pixels are dropped from
Sugarloaf::image_data, but the GPU texture stays resident for the life
of the window.

Because atlas ids are monotonic (every sixel/iTerm2 image gets a fresh
GraphicId), a long-running session that renders many images — shell
prompts with inline thumbnails, image-heavy TUIs, repeated icat — grows
image_textures without bound. The LRU only caps it at the byte budget;
below that ceiling, dead textures accumulate indefinitely.

Fix

Sweep image_textures at the start of prepare, keeping only keys that
still exist in image_data:

self.image_textures.retain(|id, _| image_data.contains_key(id));

image_data is the source of truth for what pixels exist:

• A texture is only ever inserted for a key that already has an
  image_data entry (the upload path reads from image_data).
• rio drops the image_data entry exactly when the graphic is freed
  (atlas graphic's last cell dropped, or kitty image deleted).

So a texture with no matching image_data key can never be drawn again —
it is safe to drop, and doing so is the only thing that reclaims its
VRAM. An image that is merely scrolled off-screen keeps its image_data
entry (so it can re-upload when scrolled back), and therefore keeps its
texture; only genuinely-freed textures are swept.

Cost

O(live GPU textures) per frame — and this sweep is exactly what keeps
that set small, so it is bounded by real on-screen/scrollback image
count, not by session history. No measurable per-frame overhead.

Testing

• cargo fmt --all -- --check — clean
• cargo clippy -p rioterm --all-features — no new warnings
• cargo test --features wgpu — all suites pass (0 failed)

Note

This PR is independent of, but shares files with, my per-terminal image
key change (#1752). It is branched on top of it to avoid a textual
conflict in renderer/mod.rs; the fix itself does not depend on that
change and can be rebased onto main directly if #1752 is not taken.