Detects the language of a given text using a fine-tuned BERT model. Returns the top-k most probable languages.