Rework Update Mechanism

The current learner model update mechanism (for competencies but also very much for confidences) is very ad-hoc and needs to be put on more rigorous foundations.

This will probably involve (?) reformulating the system as a bayesian network, or at least that's an option to be closely considered. We might also start to compute parameters such as how strong an update from, say, a correct answer on a certain LO for a certain learner should be on both "column-wise" (per LO) but also "row-wise" (per learner) level.

A lot of this needs a lot of design, still. People have many ideas and little concrete advice.