114 inconsistency between fitted and predict

[mahendra-mariadassou]

Fix #114 by changing behavior of predict() method (and corresponding S3-methods) for PLNfit objects:

• if newdata argument is missing, returns fitted
• if newdata is provided but not responses, returns either $EZ = XB + O$ (link type) or $EZ = exp(XB + O + 0.5\text{diag}(\Sigma))$ (response type) as the best estimates of $M_{new}$ and $S_{new}$ are then 0 and $\text{diag}(\Sigma)$
• if responses is also provided, performs one VE step to estimate $M_{new}$ and $S_{new}$ and returns either $EZ = XB + O + M_{new}$ (link type) or $EZ = exp(XB + O + M_{new} + 0.5 S_{new})$ (response type)

Open question for @jchiquet : should we return $M_{new}$ and $S_{new}$ as attributes of the predictions (like we do for predict_cond() ? I think there's not a lot of added value compared to using optimize_vestep() if one is mostly interested in M and S but you may have a different opinion.

[jchiquet]

Open question for @jchiquet : should we return and as attributes of the predictions (like we do for predict_cond() ? I think there's not a lot of added value compared to using optimize_vestep() if one is mostly interested in M and S but you may have a different opinion.

I agree with you, no need to send back M or S as attributes.