fix: Avoid O(n^2) backtracking in HTML block close and tilde interrupt regexes

[hong4rc]

Follow-up to markedjs#4013 — two more spots in src/rules.ts where a block-level regex backtracks quadratically on a long single line. Split into one commit each.

The first is the HTML block close. markedjs#3991 rewrote the close of the closing-tag, processing-instruction, declaration, and CDATA branches to [^\n]*\n+ so trailing text on the close line is kept. Keeping the trailing text is right, but the \n+ isn't: it requires at least one newline, so when the input ends without a trailing newline the close can never match and the engine retries every split point of the preceding lazy [\s\S]*? before finally falling through to $. That's O(n²):

marked.parse('<!x ' + 'a>'.repeat(80000));            // ~5s
marked.parse('<script>' + 'a</script>'.repeat(40000)); // ~6.5s

Switching \n+ back to \n* closes on the first match. Whenever a trailing newline is present (the normal case) \n* and \n+ consume exactly the same text, so output is unchanged — markedjs#3991's own trailing-text fixtures still pass.

The second is the paragraph/table/blockquote interrupt check. Its fence sub-pattern is (?:`{3,}(?=[^`\n]*\n)|~{3,})[^\n]*\n. The backtick branch is guarded by a lookahead, but ~{3,} isn't, and it overlaps with the following [^\n]*, so a long run of tildes with no newline backtracks quadratically:

marked.parse('intro\n' + '~'.repeat(80000));           // ~5s

Because ~{3,} is always immediately followed by [^\n]*, ~~~[^\n]* matches exactly the same strings, so replacing ~{3,} with ~~~ removes the overlap without changing what matches. The real fenced-code tokenizer (which captures the fence length via a group and a $ alternative) is left untouched — only the three interrupt copies change.

Both changes are behavior-preserving: the full spec and unit suites pass with identical output, and I added a quadratic_*.cjs guard next to markedjs#4013's for each.

[hong4rc]

Promoted to upstream markedjs/marked. Closing the fork-review copy; keeping the branch.