Emergence of In-Context Reinforcement Learning from Noise Distillation