Sample-EfficientReinforcementLearningIsFeasible forLinearlyRealizableMDPswithLimitedRevisiting