Detect style run breaks via cheap CellStyleKey rather than full CellStyle comparison

[emil-e]

What

The hot inner loop in buildRowContent previously called readCellStyle on
every terminal cell and compared the result field-by-field with CellStyle.eql()
across all ten style attributes (fg, bg, bold, italic, faint, underline,
underline_color, strikethrough, inverse, hyperlink).

This PR replaces that with a two-field CellStyleKey (style_id + hyperlink)
that is cheap to read and compare. The full readCellStyle call is now deferred
to run boundaries only — once per style change rather than once per cell.

style_id is Ghostty's unique identifier for a cell's complete attribute set.
hyperlink is tracked separately because Ghostty does not fold hyperlink state
into the style ID, but Ghostel uses text properties for it.

Supporting changes:

• RunInfo.style becomes ?CellStyle (null = default style), removing the
  isDefault() guard from applyStyle and making the null check explicit at
  the call site
• Wide-character spacer tails compensate with end_char -= 1 after the style
  logic runs, preventing boundary overcounting that would otherwise bleed style
  from a styled wide character onto the following default-style cell
• CellStyle.eql() is removed; std.meta.eql on CellStyleKey replaces it
• The explicit "close final run" block at the end of the loop is no longer needed

Benchmark

All runs: zig build -Doptimize=ReleaseFast, Elisp byte-compiled.
Machine: Apple M5 Pro, macOS 26.4.1, Emacs 30.2.

Streaming (chunked write + periodic redraw, no PTY)

Scenario	main	branch	Δ
stream/incr	25 ms	20 ms	−20%
stream/full	26 ms	20 ms	−23%

TUI frame rendering (full-screen rewrites)

Scenario	main	branch	Δ
incr 24×80	11 379 fps	13 466 fps	+18%
full 24×80	11 448 fps	13 805 fps	+21%
incr 40×120	5 182 fps	6 550 fps	+26%
full 40×120	5 285 fps	6 539 fps	+24%

TUI partial update (bottom-row update over static screen)

Scenario	main	branch	Δ
full 24×80	14 180 fps	18 110 fps	+28%
full 40×120	6 801 fps	9 634 fps	+42%
incr 24×80	124 031 fps	133 969 fps	+8%
incr 40×120	87 782 fps	99 775 fps	+14%

Engine micro-benchmarks (single bulk call, no PTY)

Scenario	main	branch	Δ
plain/incr	128 MB/s	135 MB/s	+5%
styled/incr	89 MB/s	95 MB/s	+7%
unicode/incr	179 MB/s	197 MB/s	+10%

Real-world PTY (cat 1 MB through process pipe)

Scenario	main	branch	Δ
plain/incr	105 MB/s	111 MB/s	+6%
plain/full	107 MB/s	110 MB/s	+3%

PTY improvements are modest: throughput there is dominated by process I/O and
timer overhead rather than per-cell render cost.

[emil-e]

Flaky test, as far as I can tell.

[dakra]

Very nice. Thanks.

BTW, if you want I can give you maintainer rights here on this repo.
Then you could trigger the flaky pipeline yourself and also close PRs/Issues etc.
But don't feel obligated to help maintain, just in case you're interested since you're also
very invested in ghostel already :)

[emil-e]

@dakra I am indeed pretty invested by now and I'd love to have a chat about. I'll drop you an e-mail!