Pretty much anything can be interpreted as a distribution or some kind of energy model, that doesn't mean that everything is the same

First of all, the softmax output of a Seq2Seq is just a formalization of the categorical distribution with which we model our problem

Attention in other hand is more complicated: to begin with, the attention module learns a residual of the input, in other word is "how should i correct the current input of the layer", so attention is actually a "smart weighted sum for the residual"

At that point, how you implement the attention, is up to you, indeed there exist alternatives to softmax attention, such as linear attention, which definitely loses the notion of "distribution", but somewhat keeps the notion of "attention" (since you put high magnitude scores to something you want, and small magnitude scores to what you don't care)

So yeah, at most you can say that attention is "a probability distribution over the tokens that allows the network to improve the current input via a residual connection"... not too sure that this is a helpful interpretation whatsoever