Breaking the MoE LLM Trilemma: Dynamic Expert Clustering with Structured Compression