Fix for measure-driven bisim partitions

[sjunges]

This PR resolves two things:

First, as raised in #91, the number of states are not always coinciding and even more, the labels are confusing (seem even wrong). The point here is that sparse bisim allows "measure-driven partitions" that merges prob0/prob1 states.

1. This PR adds an option to switch this behavior off.
2. This PR fixes the behavior for expected rewards by using the R=0 and R=inf partitions.
3. This PR makes enabling measure-driven partitions more robust.

I believe that together with #991, this may completely address the issues in #91, but this is to be tested.

[sjunges]

We will abandon measure-driven bisim in favour of goal-state merging as a preprocessing step.