Lower Bounds for Learning in Revealing POMDPs