The approach with 3 separate runs is more faithful to the modeling idea of finding differences between/among the varying stimuli. In the 1 big run approach, the separate stimuli that are glued together will get fit with 1 beta, except for the left-out individual stimulus. In the 3 run approach, the different stimuli (again, except the left-out one) will get 3 different betas, which should be slightly better.
Of course, you could try both methods, and see if your results are markedly different. I don't think they would be, since in the end you are mostly interested in the betas that come from the left-out stimuli, and the other betas are background piffle. So perhaps estimating background piffle slightly more correctly won't really matter. Perhaps.