Provably Efficient Learning in Partially Observable Contextual Bandit