Learning in POMDPs is Sample-Efficient with Hindsight Observability