Rethinking Diverse Human Preference Learning through Principal Component Analysis