feat(ffi): add normalize_sentence_lang for multi-language ITN (#83)

[Alex-Wengg]

Closes #83.

Problem

The library covered all 7 languages for single-expression ITN (normalize_with_lang) and for sentence-level TN (tn_normalize_sentence_lang), but sentence-level ITN was English-only. Non-English dictation cleanup (Parakeet STT post-processing) got no number formatting, because callers only have full sentences — not the clean single expressions normalize_with_lang expects.

Change

Adds the symmetric ITN sentence API across every binding layer:

Layer	Added
Rust lib	normalize_sentence_lang(input, lang) + normalize_sentence_with_max_span_lang(...)
C FFI	nemo_normalize_sentence_lang(input, lang) + header decl
Swift	NemoTextProcessing.normalizeSentence(_:language:)
WASM/npm	normalizeSentenceLang(input, lang)

Two tagger architectures

The interesting part is that ITN languages split into two families, dispatched in normalize_sentence_with_max_span_lang:

• en/fr/de/es — expression-level taggers that expect a clean span. These are scanned with the existing sliding window (sentence_loop) via a new parse_span_lang, which delegates to the per-language try_*_taggers helpers (already ordered by priority).
• hi/ja/zh — scanning taggers that already replace number spans in place across the whole input, so normalize_with_lang is the sentence-level entry point for them and max_span_tokens doesn't apply.
• Unrecognized codes (and "") fall back to English sentence mode.

This is why the TN side's single uniform macro couldn't just be copied — the ITN language module sets are non-uniform (e.g. ja has no whitelist/money/measure, hi adds address).

The headline behavior the issue asked for:

// whole-input normalize_with_lang leaves this untouched (not a clean expression)
normalize_sentence_lang("j'ai vingt et un ans", "fr")            // "j'ai 21 ans"
normalize_sentence_lang("j'ai vingt et un ans et trente chiens", "fr") // "j'ai 21 ans et 30 chiens"

Tests

• tests/multilang_itn_tests.rs — all 7 languages, multi-expression sentences, English/unknown fallback, empty/whitespace, max_span behavior.
• FFI unit tests (nemo_normalize_sentence_lang incl. null handling).
• Rust doctest on the public fn.
• Node WASM smoke assertion.

Full suite green (cargo test, --features ffi, --doc), cargo fmt --check clean, release+ffi build exports the symbol.

Known limitation (pre-existing, not introduced here)

The Spanish cardinal tagger over-eagerly joins "y" — normalize_with_lang("diez y veinte", "es") already returns "30" on main. Sentence-scanning makes this reachable more often (e.g. "manzanas y diez peras" drops the "y"). This is a tagger-quality issue in itn::es::cardinal, independent of the FFI wiring; worth a separate follow-up rather than expanding this PR's scope. Test assertions avoid "y"-joined Spanish cases.