Towards Unbiased and Accurate Deferral to Multiple Experts