feat(svm): draw the text run a metafile names, with the font it names

[andiwand]

🤖 Generated with Claude Code

Stage 3 of #772 and the whole of #95, stacked on #784 → #779. Review those
first; this branch's base is fix/svm-graphics-state.

Two thirds of every action in the corpus is text, and we drew almost none of
what the file says about it.

A text action names a run

MetaTextAction(point, string, index, length) — and the string it carries is
the whole paragraph. We drew the string, at every run's position.

odt/style+charset+svm-various-1.odt is the proof: its one metafile makes 29
text actions out of one Hebrew paragraph, so the reference rendering is that
paragraph printed 29 times on top of itself — an unreadable smear. With the run
applied it reads as the sentence it is, and the page's visible text loses 505
characters, every one of them a duplicate.

The run is taken in utf-16 units, before the conversion to utf-8, because that
is what the offsets count.

The font (#95)

italic, bold, underline, strikeout, and the orientation, which rotates the run
about its own start (tenths of a degree counter-clockwise, where svg turns
clockwise). Of the 7962 fonts in the corpus, 3068 are italic — every one of
them a formula variable we were drawing upright.

TEXTALIGN

It says whether the draw point is the top, the baseline or the bottom of the
run. #772 reads it as centre/right alignment for chart labels; it is not —
vcl has no horizontal text alignment, a run always starts at its point. It
still has to be written out, because its default is ALIGN_TOP and svg's is
the alphabetic baseline. svgwriter.cxx shifts the point by the font's ascent,
having the metrics; dominant-baseline leaves that to the browser, which has
them too.

The measurements the file already took

A TEXTARRAY carries a dx array — where each character ended, as the author's
font measured it — which becomes an x list, one position per character. A
STRETCHTEXT carries the advance its run has to fill, which becomes
textLength + lengthAdjust. Both are what keep a formula together when the
viewer's font is not the author's, and there are 20335 stretch texts in the
corpus.

An x list is only written when it has exactly one entry per character;
anything else is logged and dropped rather than guessed at.

Verification

8 new tests (23 in the svm suite), each built from inline bytes.

Rendered: the four svm fixtures and every changed page in headless Chrome. The
formulas of odt/Vektoranalysis Zusammenfassung.odt (1032 metafiles, 20k
stretch texts) come out italic and correctly spaced; odt/ruski.odt,
odt/arbeitsplanKW21.odt and the charts are unchanged in content and tighter
in spacing. Visible text is identical to the reference everywhere except the
Hebrew document above, where it loses the duplicates.

Needs the same reference-output regeneration as #784 — 23 files, all of
them documents carrying an svm.