kcm: first change state, then load state

the way this was before doesn't really make sense to my mind. what makes more sense is that we'd first initialize a "default" state (= no changes) and then instruct the KCM to actually load its real state. the load() may totally entail changes being made to the data that require saving.

notable example: the translations kcm may detect that some of the configured languages are no longer supported and wish to drop them from the config. in that instance load() will also produce a state change that we'd previously have forcefully undone. double bad cause this then also desynchronized the state the KCM impl thinks it has, with the actual state it has vis a vis the shell (i.e. subsequent changes never toggled the dirtyness anymore cause from the impl's POV it had already marked itself dirty)