(I don't mind if it would use only the naive approach and just multiply the probabilities. But it could be a problem if the final probabilities are very different when they are estimated with the correct method and the naive method.)
This is interesting, are they linked in some manner to national polls, or is their something more nuanced?
Somewhat intuitively, the simulations themselves have concepts of "national movements", where a candidate can (with some probability) gain 1%, 2%, etc. across all states.
The state probabilities are then derived from how many simulations/total candidate A won. However, partly because these simulations actually had national movement effects, you'll find that odds of winning both Florida and Ohio (as in # simulations/total candidate A wins both states) are not the state odds multiplied (they odds actually should be higher).
That said, with the raw simulation data, the graphic could show probabilities for every combination.
Also, we did a variation that used state-level probabilities to weight the tree. This gave a sense not just of the logical possibilities, but of the likelihood of each, which I liked. However, the FiveThirtyEight state-level probabilities are not fully independent, so you can't multiply them together to compute conditional probabilities. Perhaps next election!