BiasICL: In-Context Learning and Demographic Biases of Vision Language Models