Curve Your Attention: Mixed-Curvature Transformers for Graph Representation Learning