Allow cvrisk( , grid = 0:mstop)

[hofnerb]

As we are now (see #64) able to fit models with mstop = 0 we should be also able to

• extract a sensible risk for the offset model.
• allow cross-validation starting from 0, i.e., allow that no base-learner is selected.
• Perhaps, we have to fix predict and fitted for an offset model. Currently, a scalar is returned bu actually a constant vector of offsets might be more suitable. However, we then need to make sure that the subset function in mboost_fit works as desired.

The first point might introduce a lot of changes as we cannot assess a vectors zeroth element (hence all risks are shifted by one index). Consequently using the extractor function risk.mboost() might avoid to many changes.

[hofnerb]

cvrisk for CoxPH is currently only working for grid = 1:mstop. Atm, it changes the grid if 0 is included and issues a warning.