Deactivate a selection when terminal output repaints it - #705

A selection left active over terminal output kept its buffer positions
while the app rewrote the cells under them, so it ended up highlighting
text that was never selected. The reported case is an alternate-screen
program redrawing its whole grid after a window resize (e.g. a minibuffer
growing and shrinking the other windows).

This is xterm's keepSelection: false behaviour: output that touches the
selected rows drops the selection.

The change

The renderer already publishes the buffer range it rewrote in
ghostel--repainted-region. ghostel--deactivate-repainted-region
deactivates the region when that range overlaps it, with
select-active-regions bound off so the replaced text does not reach the
primary selection. No new state, no Zig change; it costs 54 ns on a
redraw with no region active, the second-cheapest of the eight per-redraw
calls.

Carve-outs:

• An empty region is spared. It highlights nothing, and dropping it would
  also resume window following, throwing away a scrolled-up view.
• Line mode is exempt: it passes FULL on every redraw, so its repainted
  range is always the whole buffer. A TUI taking the alternate screen
  pauses line mode to semi-char, so alt-screen selections stay covered.
• Only a redraw that actually painted is tested.
• A signalling deactivate-mark-hook is demoted so it cannot skip the
  link scan, the line-mode resume and the password check that follow.

Who sees a change

Only ghostel-mouse-drag-input-mode set to emacs or nil. The default,
copy mode, pauses redraws, so nothing can repaint under a selection
there. Under a timer-refreshing TUI (htop and friends) a selection in the
live modes now dies at the next refresh rather than lying about what it
covers; copy mode remains the way to hold one.

Testing

make -j8 all green. Four new tests: three elisp (overlap with
select-active-regions off, a boundary table of six geometries, line
mode spared, a signalling hook contained) and one native end-to-end that
drives the reported path — alt screen, select a row, resize, app repaints
through ghostel--filter. The native one fails when the new function is
stubbed out.

Verified live in a sandboxed Emacs against htop: the stale-selection
symptom reproduces on demand with the fix neutralised, and normal-screen
selections in scrollback are untouched by new shell output.

Known gaps

• Evil visual state draws its selection with evil's own overlay, so
  deactivate-mark leaves the highlight in place. Mouse selection changes after resizing an alternate-screen Ghostel window #704 is not fixed for
  evil users; that belongs in evil-ghostel.
• The published range is the union of the flushed spans, so a repaint of
  two distant rows can drop a selection sitting on untouched rows between
  them. Accepted imprecision; a per-span range would mean publishing a
  list from the renderer.

Fixes #704