ReLIC: A Recipe for 64k Steps of In-Context Reinforcement Learning for Embodied AI