Towards Understanding How Transformers Learn In-context Through a Representation Learning Lens