feat: GNOME support via a transparent click-through overlay backend

[dahui]

Summary

Fixes #16 — on GNOME the drawer came up as a small unusable box in the middle of
the screen.

The drawer anchors itself with the layer-shell protocol (zwlr_layer_shell_v1),
which is a wlroots extension, not part of wayland-protocols. KWin, Hyprland
and Sway implement it; GNOME's Mutter never has. We called InitForWindow
without a support check, and the failure is silent: the library logs one
G_LOG_LEVEL_WARNING and returns, then every SetLayer/SetAnchor/SetMargin/
SetMonitor/SetKeyboardMode warns once and no-ops. Nothing aborts. Since the
top+bottom anchors were the only thing supplying a height, the window collapsed.

This is not a library-version problem and not a "GNOME doesn't support Wayland"
problem, and no package can fix it — gtk4-layer-shell is the client side and
is already a package dependency; the protocol has to come from the compositor.

Adds a third backend selected automatically when layer-shell is unavailable: a
fullscreen transparent window with the drawer against the right edge, its input
region restricted to the drawer so everything outside is click-through and the
rest of the desktop stays usable. Core Wayland gives a client no way to position
its own window, so fullscreen is what makes edge placement possible at all.

Changes

• internal/gui/overlay — new backend (fullscreen, transparent,
  click-through) implementing the existing 5-method Backend interface.
• internal/panelgeom — new pure package for the panel rectangle and slide
  interpolation, with invariant tests. Picked up automatically by PURE_PKGS.
• gui.go — three-branch backend selection. layerShellUsable() checks the
  GDK backend before calling IsSupported(), which asserts on a non-Wayland
  display and would otherwise log assertion 'GDK_IS_WAYLAND_DISPLAY' failed at
  CRITICAL in every X11 session.
• layershell.go — stop logging backend mode=layer-shell unconditionally.
  It reported success even when init had failed, which is the single most
  misleading line in the diagnostic path and why Creates a small window at the center of the screen that cannot be interacted with. #16 looked like a rendering bug.
• controls.go / tdp.go — three identical ScrolledWindow setups replaced
  by one newDrawerScroll() with SetMinContentHeight. A GtkScrolledWindow
  reports a minimum height of 0 and SetVExpand only distributes surplus, so
  with nothing sizing the window from outside the drawer collapses. Deliberately
  not SetPropagateNaturalHeight(true) — that trades no default size for a bad
  one (the custom view would request more than the screen height).
• Docs — a GNOME support section (why the workaround exists, how it works,
  what is degraded), plus troubleshooting for journalctl --user: ours is a
  systemd user unit, and sudo journalctl -u z13gui prints -- No entries --,
  which reads as though the program never ran. That is what sent the reporter
  down the wrong path.
• CLAUDE.md — records the "never call into gtk4-layer-shell without checking
  IsSupported()" rule and the new backend.

Testing

make build, make fmt-check, make lint (0 issues), make test (11 packages),
make race all pass.

Verified on hardware (KDE, using GDK_BACKEND=x11 to force the same
"no layer surface" path GNOME takes):

• Backend selection both ways: GDK_BACKEND=x11 → backend mode=overlay;
  native Wayland → backend mode=layer-shell protocolVersion=4, unchanged.
• Geometry panel={Y:80 H:1440} on a 2560x1600 output — 5% margins, full height.
• Positioning fix confirmed by measurement. GtkFixed allocates children
  their natural size and SetSizeRequest is only a minimum, so the panel came
  out 343px wide against a 320px request (CSS border/padding) and hung 23px past
  the screen edge, clipping the second column of every row. Positioning now
  derives from the allocated width: x=2217 w=343 → right edge exactly 2560.
• Focus-loss dismiss fired on its own (overlay: dismissing on focus loss).
• No new GTK warnings: the two GtkGizmo slider warnings appear identically on a
  pre-change baseline binary, so they are pre-existing.
• The panelgeom tests fail when the width clamp is mutated, i.e. not vacuous.
• mkdocs build --strict passes; verified id="gnome-support" exists in the
  generated HTML and all five cross-links resolve.

Not verified — needs a real GNOME session. KWin is not Mutter:

• Click-through. The call succeeds without error, but xdotool getmouselocation
  resolves to a compositor container under XWayland and xwininfo does not expose
  the input shape, so neither could confirm it.
• Whether Mutter hides the top bar while the drawer is open, and whether it
  appears in Alt-Tab. Both are documented as "may". If the top bar behaviour is
  objectionable the fallback is a monitor-sized non-fullscreen window — at exactly
  monitor size the compositor's own placement lands it correctly, so positioning
  is still not needed. One-line change.
• Scaling. Under GDK_BACKEND=x11 GTK reports scale 1, so the drawer renders at
  343 physical px and looks small; on GNOME Wayland it should get the compositor
  scale natively.

Checklist

• make build passes
• make lint passes
• go test ./internal/theme/ -v passes
• Documentation updated if needed