Mitigating Label Biases for In-context Learning