Neural Active Learning Beyond Bandits