fix: case fold reference link labels - #4077

Description

Reference link labels are matched with toLowerCase(), but CommonMark asks for a Unicode case fold:

To normalize a label, strip off the opening and closing brackets, perform the Unicode case fold, strip leading and trailing spaces, tabs, and line endings, and collapse consecutive internal spaces, tabs, and line endings to a single space.

toLowerCase() doesn't get there, so labels that only agree once folded never match:

[ẞ] and [Straße].

[SS]: /sharp-s
[STRASSE]: /street

renders as literal text today, because 'ẞ'.toLowerCase() is 'ß' — never 'ss'. Round-tripping through upper case reaches the fold, which is what commonmark.js does in normalizeReference():

.toLowerCase()
.toUpperCase();

I lowercase once more afterwards so def.tag and the tokens.links keys stay in the lower case form they've always had — an uppercase key would be a visible change for anyone reading lexer.tokens.links or a def token, and it isn't needed to get the fold.

The two link-map lookups in Lexer (linkInText and the reflink mask) used the label verbatim rather than normalized, so they go through the same helper. Without that they'd start missing definitions they used to find as soon as the stored keys are folded. As a side effect the mask now also handles labels whose case differs from the definition.

This enables CommonMark example 540, taking Links from 81/90 to 82/90.

Contributor

• Test(s) exist to ensure functionality and minimize regression (if no tests added, list tests covering this PR)
• Docs have been added/updated (for bug fixes/features)

Verification notes:

• Removed the shouldFail flag from example 540 in both test/specs/commonmark and test/specs/gfm.
• Added test/specs/new/reflink_case_folding.{md,html} covering ẞ/SS, Straße/STRASSE, Dž/DŽ and fi/FI. I generated the expected HTML from commonmark.js 0.32 and it is byte-identical to marked's output with this change; the fixture fails on master.
• npm test is green: specs 1789 -> 1791 passing, unit 191 passing, 0 failures, lint clean. No existing test needed adjusting.