Efficient Reinforcement Learning with Impaired Observability: Learning to Act with Delayed and Missing State Observations