On the Effectiveness of ASR Representations in Real-world Noisy Speech Emotion Recognition