Can Compressed LLMs Truly Act? An Empirical Evaluation of Agentic Capabilities in LLM Compression