Break on decreasing LL

[simon-hirsch]

We should allow breaking if the model gets worse. Currently we only check if the LL changes (better or worse).

There is a simple fix (i.e. just break) but the better solution would be to actually go back one step and recover the optimal model. Open questions to resolve this properly:

• In the inner iteration, we can just "go back" one step.
• In the outer iteration, we can go back on step or one distribution parameter.

Reported by Katarzyna Chec.