Oracle-free Reinforcement Learning in Mean-Field Games along a Single Sample Path