Does RL Incentivize Reasoning Capacity in LLMs Beyond the Base Model?