Analyzing LLM Behavior in Dialogue Summarization: Unveiling Circumstantial Hallucination Trends