Positional Attention: Out-of-Distribution Generalization and Expressivity for Neural Algorithmic Reasoning