Model families, component inference, and individual crossvalidation

We added the construction of model families and component inference using model averaging.
The first version of within-subject crossvalidation is also provided, but using only standard encoding-model evaluation criteria (R and R^2).

Optimisation stability, Component models

• increased stability of NR optimization
• inference on pattern components in model families

Newton-Raphson optimisation

Changed from conjugate gradients to Newton-Raphson for small to medium scale problems.

Initial release of full toolbox

Initial release for OHBM2017.
The recipes and basic methods should be fully functional.
Optimization of models is purely achieved by gradient descent (pcm_minimize).
The next version will use faster algorithms adaptively whenever possible.