Use a linear predict when possible

[janosg]

We use an unscented Kalman predict even if all transition equations are linear. However a linear predict step would be much faster and the predict step is the bottleneck in most models.

The linear predict step is already implemented and tested. The only thing that's left is to integrate it into the likelihood function. Among others this will need the following steps:

• write a function that determines if the linear predict can be used
• construct the transition matrix from the parameters
• construct other input arrays
• select the correct predict function