In-Context Learning (and Unlearning) of Length Biases