Reducing Planning Complexity of General Reinforcement Learning with Non-Markovian Abstractions