Softmax as Linear Attention in the Large-Prompt Regime: a Measure-based Perspective