Mitigating Copy Bias in In-Context Learning through Neuron Pruning