Loocv updates

[jelmerborst]

Loocv:

• loocv consolidation to handle backward models, single models and condition based models
• new loocv_func that does loocv correctly: it applies a provided func to the data of n-1 subjects, and then calculates the likelihood on the left out subject. This function can be as complex as you want and also test multiple models at once. It's also not much slower in the end.

Cross collleration:

• fixed using numpy crosscorr method. Works now independent of template width and shape

In addition, while doing these a bunch of small fixes which only happen in edge cases, in particular:

• remove nans on pmfs
• min_iteration >= 1 instead of >
• checking for -inf

[jelmerborst]

Fixed based on your comments above.