Fontique: 3-tier font fallback

[nicoburns]

• Fixes Fontique fails to find font for ❯ char (U+276F) on macOS #678
• Fixes Font Fallback not working for at least 汉 on Windows #644

Generated with Fable High.

Motivation

Font fallback completely fails to find a matching font for some glyphs.

This is because font fallback was keyed on (script, locale), and resolved by asking the OS for a single font family which matches a per-script sample string. If that one font lacked the actual character (for example, CJK extensions, symbols, punctuation shared between scripts, scripts without samples) the Fontique gave up, even if there was actually an installed font that had the required glyph.

Changes made

Implement 3-tier fallback:

Tier A (all backends) - Ordered list of fonts per script (script, locale)

This is the existing fallback mechanism, but modified to work with a list of fonts rather a single font per (script, locale).

• CoreText uses CTFontCopyDefaultCascadeListForLanguages
• fontconfig uses font_sort with coverage trimming instead of font_match
• Android uses all matching locale/script chain entries instead of the first

Tier B (skipped on Android) - per-character system fallback

• Added Query::set_fallback_chars() where Parley passes the specific characters we are falling back for
• When script/locale lists (Tier A) lack coverage, the backend is asked for a font covering the actual text (CTFontCreateForString / IDWriteFontFallback::MapCharacters / FC_CHARSET match)
• This is cached per (character-set, locale) in the collection.

Tier C (all backends) - exhaustive coverage scan

As a last resort, all families are checked for real cmap coverage (guaranteeing a glyph is found if any font has one)

• Fonts are sorted by name, with hidden/.-fonts last, and LastResort excluded)
• Positive and negative results are cached, so the scan happens once per unique character set.
• This scan is expensive, so there is an opt-out via a new exhaustive_fallback: bool runtime setting on Collection

This PR also:

• Removes the "hani hack" (macOS: CJK punctuation marks not rendered correctly #597), which is now covered by the general per-character fallback mechanism
• Fixes a latent parley bug where empty layouts shaped their cursor-metrics run with a zero font size.

Verification Performed

The AI has run temporary integration tests (not included in this PR) on macOS and Ubuntu.

Tested in Blitz on macOS, and it definitely seems to work and find glyphs where main doesn't.

[nicoburns]

Is 8 chars a sensible size here?

[nicoburns]

I think we can eliminate this allocation.

[dfrg]

I don’t have time for a deep dive on this right now but, at a glance, I’m generally supportive of this with a couple caveats:

1. The char lookup and cache is iffy to me. I’m not opposed in principle but I think we might be able to get by with appending a list of known large coverage fonts to the fallback chain for every query.
2. I’d like to avoid adding an exhaustive search, even as an option. This is a perf killer and pretty much always produces the wrong result.

Big +1 to returning multiple fallback fonts from the system. We should probably also extend the script sample array to support multiple samples per script. This would help, for example, with selecting the -ExtB CJK fonts on Windows.