Fit-to-screen image scaling (v0.3.0)

[whallil]

Summary

Adds fit-to-screen scaling so images (and Mermaid diagrams) larger than the terminal are downscaled to fit instead of overflowing onto following text.

• Replaces image_rows() (reserve-only, height capped at terminal height → oversized images overflowed) with placement_cells(), which returns a concrete (cols, rows) box.
• show.py passes that box as Kitty's c=/r= placement keys, so Kitty does the resampling, preserving aspect ratio. Images that already fit render pixel-for-pixel.
• Vertical fit reserves DRIFT_MARGIN + 2 rows of headroom so a downscaled image plus its anchor offset stay on one screen (no clipped bottom).
• Drops the leading PTY newline (was an uncommitted top gap).
• Applies to Mermaid too, since mermaid.py displays through show.py.

Docs

• SKILL.md: documents fit-to-screen; replaces the obsolete "taller than screen overflows / not implemented" caveat; adds an honest "vertical placement is approximate" note; adds guidance to keep inline Mermaid diagrams compact (graph LR, few nodes) since tall ones get scaled and can wash out.
• README.md "Known limitations" updated.
• CHANGELOG.md [0.3.0]; plugin.json → 0.3.0.

Testing

• placement_cells() verified both branches: small 240×120 → native (24,6) (no scaling); tall 400×2970 → scaled (11,36), aspect preserved.
• End-to-end renders confirmed in a live Kitty session: a photo that previously overflowed (43 unscaled rows) now scales to 36 and fits; Mermaid diagram rendered locally (mermaid-cli + Chrome) and displayed.
• Visual placement confirmed by the maintainer.

Note

The Unicode-placeholder placement approach explored in a worktree was abandoned: the editor filters U+10EEEE (PUA) characters out of stdout, so it never renders. This PR keeps the working direct-paint path and grafts on only the resize logic. Worktree purged.