NoisyTwins: Class-Consistent and Diverse Image Generation through StyleGANs