Use model context in model.logpt

[aseyboldt]

Since theano doesn't compute test values outside the model context anymore, we need to do all model.logpt computations inside the model context. This lead to errors in the getting_started notebook, when it used find_MAP outside of the context. This used model.logpt which got cached without test values. This led to an exception in a later sampling statement.

[junpenglao]

LGTM