APTBench: Benchmarking Agentic Potential of Base LLMs During Pre-Training