Maximal Multiverse Learning for Promoting Cross-Task Generalization of Fine-Tuned Language Models