Semantic Image Synthesis via Class-Adaptive Cross-Attention