Dyna-LfLH: Learning Agile Navigation in Dynamic Environments from Learned Hallucination