On Statistical Discrimination as a Failure of Social Learning: A Multi-Armed Bandit Approach