Robustness of Speech Separation Models for Similar-pitch Speakers