Custom Algorithm-based Fault Tolerance for Attention Layers in Transformers