Text this: Zero-shot personalization of speech foundation models for depressed mood monitoring