fix: isExactMatch: strip diacritics for accent-aware exact matching

[YannBirba]

This PR was created by human (me), but all the code LLM generated.
I hope it's okay for you.

Problem

When the GGML language model suggests an accented word (e.g. Garçon) for a
user who typed an unaccented ASCII input (garcon), the exact-match scoring
rule at org_futo_inputmethod_latin_xlm_LanguageModel.cpp:1273-1286 fires
incorrectly:

// Exact match rule
bool hasExactMatch = false;
for(const auto &result : results) {
    if(isExactMatch(result.second, partialWordString)) {
        hasExactMatch = true;
    }
}
if(hasExactMatch){
    for(auto &result : results) {
        if(!isExactMatch(result.second, partialWordString)) {
            result.first -= 1.0f;  // <-- penalizes accented candidates
        }
    }
}

isExactMatch previously only lowercased and stripped ', -, and spaces.
UTF-8 accented characters (é, ç, ê, …) were compared byte-for-byte,
so "Garçon" never matched "garcon". The -1.0 penalty pushed accented
candidates below the unaccented form, defeating the LM's accent predictions.

Root cause

isExactMatch did not normalize Unicode diacritics to their ASCII base forms
before comparison. A user typing garcon on an ASCII keyboard would always
see Garcon instead of Garçon, even when the LM correctly predicted the
accented form.

Fix

Add append_diacritic_normalized(), a lightweight UTF-8 normalizer that maps
accented characters to their base ASCII equivalents:

• Latin-1 Supplement (U+00C0-U+00FF, 0xC3 prefix): all common
  European accents (French, Spanish, Italian, Portuguese, German, Nordic)
  — including uppercase variants (À-Ÿ) and ligatures (Æ→"ae").
• Latin Extended-A: Œ/œ (0xC5 0x92/0x93) → "oe".

Unmapped sequences (Ð, Þ, ß, …) keep their original bytes, preserving
stock matching behavior.

Testing

Added native/jni/tests/test_isexactmatch.cpp — a standalone test
(compilable with g++ -std=c++17) covering:

Category	Examples
French lowercase accents	garçon↔garcon, café↔cafe, être↔etre
Mixed case	Garçon↔GARCON, CAFÉ↔CAFE
Ligatures	cœur↔coeur, Œuvre↔Oeuvre
Separators stripped	l'homme↔lhomme, Hello-World↔helloworld
Negative cases	maison≠maisonn, chat≠chien
Unmapped sequences	ðtest≠dtest (preserves stock behavior)

33/33 tests passed

Also tested on-device (Android emulator, AZERTY layout) with the custom
GGML model:

Typed	Expected	Result
garcon ␣	Garçon	✓
Xafe ␣	Café	✓
bonjor ␣	Bonjour	✓
etre ␣	Être	✓
francais ␣	Français	✓

Scope & limitations

• Covers Latin-1 Supplement + Œ/œ. Other scripts (Cyrillic, Arabic, CJK)
  are untouched and preserve stock byte-for-byte comparison.
• ñ normalizes to n for exact-match purposes (intentional: beam search
  should not penalize ñ when user typed n).
• Trade-off: words differing only by accent (a/à, ou/où) will both count
  as exact matches. The LM probability still decides ranking; auto-correction
  thresholds guard unintended replacements.

[futo-cla-pr-labler]

Please sign our contributor license agreement at https://cla.futo.tech

[futo-cla-pr-labler]

Please sign our contributor license agreement at https://cla.futo.tech

[YannBirba]

Just signed the CLA