fix: pass language parameter to TTS models

[jannone]

Summary

• Both backends silently dropped the language parameter — it was accepted by generate() but never forwarded to the underlying Qwen3-TTS model, causing it to default to "Auto" detection which frequently confuses Portuguese for Spanish
• Added LANGUAGE_CODE_TO_NAME mapping (ISO 639-1 → full language name) to both PyTorch and MLX backends, matching the format Qwen3-TTS expects (e.g. "pt" → "portuguese")
• PyTorch backend: passes language= to generate_voice_clone()
• MLX backend: passes lang_code= to all 4 model.generate() call sites
• Frontend: auto-syncs the generation form's language dropdown with the selected voice profile's language via a new useEffect

Test plan

• Select a Portuguese voice profile → language dropdown should auto-sync to Portuguese
• Generate speech in Portuguese → output should sound Portuguese, not Spanish
• Switch to English profile → language dropdown syncs to English, generation works normally
• Test with other supported languages (Chinese, Japanese, etc.)
• Check backend logs for any language-related errors

🤖 Generated with Claude Code

[jamiepine]

Fixed on main in ca74c15 — both backends were silently dropping the language parameter. Applied the same approach from this PR: LANGUAGE_CODE_TO_NAME mapping + forwarding to both PyTorch (language=) and MLX (lang_code=) backends, plus the frontend profile-language sync. This was indeed the root cause of the language complaints. Thanks for the original diagnosis!