ui: add span fit_mode to split one image across outputs

[josephdunn]

Summary

Adds fit_mode = span, which splits a single wallpaper across every output it
is assigned to so the image is continuous across the physical monitor
arrangement, instead of repeating the whole image on each output (which is what
cover/contain/fill/tile do today).

Optional sub-modes control how the source is fit onto the combined canvas:

• span / span:cover (default) — fill the canvas, crop overflow
• span:contain — fit the whole image, letterbox
• span:stretch — stretch to the canvas aspect

Example

wallpaper {
    monitor =
    path = /path/to/ultrawide.png
    fit_mode = span
}

How it works

• Outputs resolving to the same wallpaper entry form a span group.
• The monitor layout is queried from Hyprland over IPC (j/monitors).
• The image is mapped onto the group's bounding box in logical layout
  coordinates (so mixed-DPI monitors stay continuous in physical size), and
  each output is handed the sub-rectangle its layout rect covers.
• Slices are rendered with cairo and cached as PNGs under $XDG_RUNTIME_DIR.
• A group of one output is equivalent to the matching plain fit mode.

Fallback

If the layout can't be read (not under Hyprland, or an output's geometry can't
be resolved), each output falls back to cover with a warning.

Dependencies

Adds cairo, hyprgraphics, libpng, and jsoncpp (CMake + nix).
cairo/hyprgraphics were already transitively present via hyprtoolkit;
libpng and jsoncpp are new.

Testing

• Verified on a 3-monitor layout (mixed transforms and resolutions): slices are
  geometrically contiguous across seams, and mixed-height/DPI monitors sample
  the correct region.
• Verified span:cover cropping, the single-output degenerate case, and the
  no-Hyprland cover fallback.

Known limitations

• An in-place monitor geometry change (resolution/scale/reposition without
  an unplug) doesn't re-slice until the wallpaper is re-set — hyprtoolkit
  exposes no "output changed" signal. Hotplug (add/remove) is handled.
• Rotated outputs work (logical dimensions handled); flipped transforms (4–7)
  are untested.
• span is config-file only; the IPC/hyprctl fit-mode enum isn't extended
  (that would be a protocol change).