feat(text): make text layout style-aware so rich text wraps correctly

[wouterlucas]

Part 3 of 4. Base is #908 — review that first; this diff is isolated against it.

Problem

MeasureTextFn had no style parameter, so line breaking measured every substring with the base font. A bold run that fits under regular metrics but overflows in bold wrapped in the wrong place, and a line reported a width narrower than its drawn extent.

PR 2 corrected the drawn extent after the fact. This corrects the measurement the wrapping decision is based on.

Approach

MeasureTextFn gains an optional start: the absolute index of the substring within the text passed to mapTextLayout. The layout engine threads it through wrapLine, the three wrap strategies, and the per-character scans in truncateLineEnd, truncateWord and splitWord.

The parameter is optional and a negative value means "not locatable in the source", which is what separators and the overflow suffix pass. The plain text path and existing callers are unaffected.

The layout engine does not learn about rich text. Each renderer supplies a measure function closed over its own parsed spans:

• Canvas switches the shared measure font per style run.
• SDF applies the synthetic bold advance and italic overhang.

Splitting at span boundaries preserves kerning within a run and drops it across one, where it does not apply.

Note on the retained width correction

PR 2 introduced correctStyledLineWidths in both renderers. It is kept rather than deleted, because a line width is still summed from word widths plus base-font separator widths, and a line-final italic overhang sits outside any word — so the accumulated width is not exactly the drawn extent.

It now just calls the same style-aware measure function instead of duplicating the span walk, which removes roughly 100 lines of near-duplicate logic across the two renderers.

Reviewer caution

Measurement is no longer a pure function of the substring — it depends on the offset too. Any future memoisation must key on (start, len), not the string alone. Called out in the commit message.

Testing

tsc --noEmit clean, 331 tests pass, eslint 0 errors. Adds 6 tests asserting that every located substring genuinely lives at the offset reported, that the overflow suffix is measured unstyled, that a styled run forces an earlier break, that a styled run off the line does not perturb it, that styled lines report greater width, and that mid-word break-all splits respect styled characters.