Multiple Choice Learning for Efficient Speech Separation with Many Speakers