Act-Then-Measure: Reinforcement Learning for Partially Observable Environments with Active Measuring