On the Self-awareness of Large Reasoning Models' Capability Boundaries