Fix #164: Runtime performance improvements

[JasonPierce]

Fixes #164

• replaceProfanity — O(n) instead of O(n*m): Replaced repeated slice + concat (each creating a full copy of the result string) with an array of parts joined once at the end. Also eliminated the offset tracking since we no longer mutate a running string.
• Vowel regexes hoisted to module-level constants: FIRST_VOWEL_RE and ALL_VOWELS_RE are created once instead of new RegExp(...) on every matched profane word during censoring.
• Word-boundary regexes hoisted to module-level constants: UNICODE_WORD_CHAR_RE and ASCII_WORD_CHAR_RE replace the isWordChar closure that was recreated on every isWhitelisted call, with a new regex instance per invocation.
• getRegex reduced allocations: Replaced [...new Set(languages.map(...))] + .toSorted() + .flatMap() + .filter() chain with a single-pass dedup loop, in-place sort, and a simple for...of with push — fewer intermediate arrays and closures.