Transformers as Multi-Task Learners: Decoupling Features in Hidden Markov Models