Learning to Navigate Under Imperfect Perception: Conformalised Segmentation for Safe Reinforcement Learning