refactor: cache language mappings and use model.device

Cache _build_language_mappings() with @lru_cache since the result
never changes. Use model.device instead of calling _get_device()
again in translate().

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

app.py

@@ -38,6 +38,7 @@ def _load_model() -> AutoModelForSeq2SeqLM:
     return AutoModelForSeq2SeqLM.from_pretrained(MODEL_NAME, torch_dtype=dtype).to(device)
 
 
+@lru_cache(maxsize=1)
 def _build_language_mappings() -> tuple[dict[str, str], list[str]]:
     tokenizer = _load_tokenizer()
     vocab = tokenizer.get_vocab()
@@ -49,7 +50,7 @@ def _build_language_mappings() -> tuple[dict[str, str], list[str]]:
 def translate(text: str, target_language_name: str) -> str:
     tokenizer = _load_tokenizer()
     model = _load_model()
-    device = _get_device()
+    device = model.device
 
     name_to_code, _ = _build_language_mappings()
     target_code = name_to_code.get(target_language_name)