Understanding Community Bias Amplification in Graph Representation Learning