Deep Language Networks: Joint Prompt Training of Stacked LLMs using Variational Inference