fix(translation): check active provider's token in translateAsync AI gate

[foegra]

Summary

Fixes #459, refs #456.

translateAsync() gated the built-in-AI translation fallback on ProofreadService.getApiKey(), which reads only the Gemini key (KEY_API_KEY = "gemini_api_key", ProofreadService.kt:645). With provider OPENAI (custom OpenAI-compatible endpoint) or GROQ, that key is empty even when the provider is fully configured, so hasAiConfigured was false and every AI-fallback branch (lines ~414/436/449/462) short-circuited to Result.failure("Translation plugin not available") with a misleading "Offline model not downloaded" toast — before any network request was made.

Proofreading was unaffected because performAsyncOperation() performs a provider-aware check (getGroqToken() for GROQ, getHuggingFaceToken() for OPENAI). GGUF translation on the Offline flavor checks the local model path instead. The gate looks like a leftover from when Gemini was the only cloud provider.

This PR changes the gate in translateAsync() to check the active provider's token, mirroring the pattern already used by performAsyncOperation():

val hasAiConfigured = when (service.getProvider()) {
    ProofreadService.AIProvider.GEMINI -> !service.getApiKey().isNullOrBlank()
    ProofreadService.AIProvider.GROQ -> !service.getGroqToken().isNullOrBlank()
    ProofreadService.AIProvider.OPENAI -> !service.getHuggingFaceToken().isNullOrBlank()
}

Verification

• Reproduced on v4.1.7 (standardfull), Google Pixel 10a / Android 16, provider = OpenAI-compatible custom endpoint (OpenAI /v1/chat/completions format), where the same provider+model works for proofreading:
  • Translation mode = Cloud/Local AI → toast "Offline model not downloaded", no request sent.
  • Workaround (dummy Gemini API key, provider unchanged) → translation succeeds through the OpenAI-compatible endpoint, confirming the gate was the only blocker.
• Root cause verified against master (≈ v4.1.7): ProofreadHelper.kt:386 + ProofreadService.kt:79/645.
• The change is a drop-in replacement at a single call site; getGroqToken() and getHuggingFaceToken() are already public methods on ProofreadService used by performAsyncOperation().

Notes

Two related cosmetic issues observed on this path (not addressed here to keep the PR minimal):

• The AI-fallback toast (translation_plugin_fallback_to_ai, "Plugin unavailable, switching to AI") shows on every translation when the plugin isn't installed, even when the user is already in AI mode.
• The failure toast says "Offline model not downloaded" even when an offline model was never involved.

Happy to follow up on those separately if useful.

Fixes #459
Refs #456