iQRL -- Implicitly Quantized Representations for Sample-efficient Reinforcement Learning