Linkify wrapped paths anywhere in scrollback, not just the viewport - #587

Fixes #582.

Paths that the terminal soft-wraps across two rows were not clickable and not
even mouse-highlighted in rails test output. Two independent causes, one
commit each.

Scan coverage

ghostel--schedule-link-detection queued only the live viewport — point-max
minus ghostel--term-rows lines. Between two queue events a flood materializes
far more rows than that, and everything above the new viewport was never
scanned by any cycle, because nothing rescans. Those rows carry no help-echo,
mouse-face or keymap, which is exactly the reported "not clickable, doesn't
even highlight" — and it also defeats the ghostel-previous-hyperlink
workaround from the issue thread, since that searches for the help-echo a
scan was supposed to have applied.

The renderer already knows which buffer range it rewrote, so it publishes that
instead. Every character of terminal output is inserted by flushSpan, which
makes coverage gapless by construction. It also closes the cursor-line skip for
free — isRowDirty repaints the row the cursor left, so a match skipped as
active input is republished and rescanned — and it is cheaper than before when
output is slow: a two-row repaint queues two rows, not the whole viewport.

Supporting changes the new region size forces:

• The queued bounds become markers. evictScrollback deletes from the buffer
  start on almost every redraw during a sustained flood, which rots integer
  bounds queued 0.1 s earlier.
• The scan drains at most ghostel--link-detection-chunk characters per tick,
  taken from the newest end, so one redraw materializing megabytes no longer
  blocks Emacs for as long as scanning all of it takes, and rows on screen are
  linkified before the backlog. Re-arming always goes through a timer, so
  ghostel-plain-link-detection-delay of 0 cannot recurse through it.
• Process exit no longer cancels a partly drained scan: the program that
  flooded has exited, so nothing will ever repaint that text again.

Row limit

ghostel--wrap-joined-region counted joined rows against
ghostel--soft-wrap-row-limit cumulatively across the whole scan region
instead of restarting at each hard newline, so a region holding more wrapped
lines than the limit stopped joining at every 50th one and lost that line's
link. Harmless while scans covered ~34 rows; load-bearing once a scan covers
whatever the renderer repainted.

Testing

make -j8 all passes at both commits, so the first is not a broken
intermediate.

New elisp tests (no module): per-logical-line row limit, chunked drain reaching
the whole pending range, bounds surviving front deletion, zero delay re-arming
through a timer instead of recursing. New native tests: ghostel--repainted-region
publication, a regression test writing far more than term-rows lines in one
write, and pending-wrap completion across two redraws. Each new test was
confirmed to fail against the pre-fix code.

Verified live under elate as well — bash in a ghostel buffer, cat of a
2000-line log with 200 wrapped 121-character paths. Post-fix: 200/200 linkified
across three runs, every link carrying the full joined target, and
ghostel-previous-hyperlink reaching all 200 from point-max. Isolating the
two bugs against that harness gives 3 misses per run at identical offsets for
the row-limit bug, and 192–193 of 200 missing for the coverage gap at
ghostel-plain-link-detection-delay 0. Revealing a buffer that flooded while
hidden scans the visible rows on the first tick and walks the backlog down over
the next seven.

Known follow-ups, not addressed here

• Line mode passes full=t on every redraw (to rebuild the prompt row after the
  input snapshot), which erases and re-renders the entire scrollback — 3.1 ms at
  a saturated default scrollback versus 0.005 ms incremental. That now queues the
  whole buffer for scanning too. It is correct rather than wasteful, since a full
  redraw wipes every detected link (measured: 40 → 0), which the old viewport-only
  scan never restored outside the visible rows — but the root fix is to stop
  forcing full redraws in line mode.
• ghostel-compile--commit-pending-frame publishes a region nothing consumes, so
  a compile buffer's final frame is never link-scanned. Pre-existing.