Keep line-mode point and markers put across a resize - #702

Fixes #699.

In line mode, after seq 1000, the first window resize threw point into
the scrollback and the next one threw it to the top of the buffer. Two
independent defects, both reproduced and measured in live sessions.

Point relocated to the middle of the window

The line-count branch of ghostel--anchor-window walked back from
point-max by the full screen-line count. That is correct only when
point-max begins a trailing empty row, which is what the renderer
normally leaves. Line mode trims that tail, so point-max ends at the
end of the prompt row, the walk-back lands one line too high, and the
prompt row falls below the window bottom. set-window-start sets
force_start, and redisplay answers an invisible point by moving it to
the middle of the window rather than scrolling.

Measured in a 22-row window with the prompt on buffer line 1002:

Graphical frames use the exact pixel anchor and were unaffected. The
fallback runs on terminal frames, on Emacs 28, and whenever the pixel
measurement is unavailable.

Point, window start and mark collapsing to position 1

The renderer pins saved markers from inside commitResize, but a FULL
redraw erases the buffer before that. bufferPosToPin derives a screen
row with line-number-at-pos, so against an empty buffer every saved
position maps to row 0 and is restored as buffer position 1. Line mode
passes FULL on every redraw, so any resize hit it. Graphical frames were
affected too.

With the mark parked in the scrollback and a width resize applied:

Point usually survived because ghostel--line-mode-restore puts it back
from its input-region offset. Once point is outside that region, which
is exactly where the first defect left it, nothing restores it.

Known ceiling

Skipping the pin restores raw buffer offsets, so a marker on a row that
rewraps during a width resize drifts by the rewrap delta instead of
following its row. Pinning accurately would mean moving the full-redraw
clear below the pin, which starts running scrollback eviction on a path
where it currently never executes. Eviction ordering here has produced
two silent desyncs before, so that trade is not worth an occasional
few-character drift. The reported collapse to the buffer start is gone.

Tests

Two native regression tests, both confirmed to fail with the fixes
reverted and the module and byte-code rebuilt:

• a full redraw carrying a pending width resize, the combination line
  mode produces and the one case the existing position-preservation
  family did not cover
• anchoring a buffer whose last row has no trailing newline

make -j8 all passes from cleared test stamps. Both symptoms were also
re-checked live, on terminal and graphical frames.