Ghost Policies: A New Paradigm for Understanding and Learning from Failure in Deep Reinforcement Learning