feat: select semantic text typefaces

[Happypig375]

Summary

• add immutable, validated semantic typeface descriptors for family, weight, upright/italic/slanted posture, advertised OpenType features, and configured small-cap glyph maps
• preserve exact local face identity and one coherent per-render snapshot, exhausting deterministic local fallbacks before globals
• keep slanted distinct from italic; unavailable slanted faces fall back to upright rather than italic/math-italic
• select a configured true small-cap glyph without rewriting source Unicode, preserving scalar enumeration and the materialized glyph run used for both measurement and drawing
• retain legacy IEnumerable<Typeface> behavior from Honor styles in local text typefaces #284 and descriptor state through absolute/relative size changes
• document native feature metadata versus configured glyph-map behavior; Typography's current GSUB path is not modified

Testing

• SemanticTypefaceDescriptorTests: 10 passed, including Source Serif Pro a → genuine A.sc glyph 1108 and its distinct advance
• TestStyledLocalTypefaces: 6 passed
• TextLaTeXParserTests: 241 passed; TextStyleTests: 17 passed
• XAML Maui/Avalonia live typeface tests: 1/1 each
• Skia FontStyles: 19/19 for both finite/infinite centered layouts
• Skia VectorProjection: 92/92 display and 92/92 inline
• Avalonia FontStyles: 19/19 across all six alignment/width variants
• all four affected Release builds passed; scoped format and diff checks passed
• no golden images changed; independent architecture review approved

Stack

This PR is based on issue-306-semantic-text-style / #331 and merges the styled-local-typeface prerequisite from #284. GitHub issue dependencies record both #306 and #142 as blockers.

Closes #307

CI hardening

A transient Windows failure exposed concurrent lazy mutation in the vendored cmap lookup while semantic font selection exercised shared typefaces. Glyph lookup is now serialized per Typeface through a weak-keyed lock, covering every reachable cmap call without globally serializing unrelated faces.

The regression test starts eight long-running workers behind a barrier, uses distinct code points to force concurrent cache growth, retains and validates all 2,048 observations, and has a five-second linked deadline. It passed three independent focused runs; the previously failing AngouriMath modulo class also passed.