scheduling transitions, not actions

In order to avoid that a scheduler (as exported currently from mcsta) may induce nondeterminism (since it currently maps to actions), it appears more handy to instead map to pairs of action and target distribution, so as to enable the reconstruction of the exact transition taken.