Noise Injection Reveals Hidden Capabilities of Sandbagging Language Models