Are Large Language Models Dynamic Treatment Planners? An In Silico Study from a Prior Knowledge Injection Angle