Are LLMs Ready for Scientific Discovery? A Capability-Oriented Benchmark for AI Scientists