Inconsistent and unclear loss calculation for contextual bandits

[jackgerrits]

The loss calculation for CB reductions is not consistent and not well documented. The current situation is:

• cb_adf records loss as calculated by an IPS estimator, except for if CB type DR or DM is in use, in which case it is using the DR estimator for loss.
• cb_explore_adf always uses an IPS estimate (including for DR and DM)

The proposed solution is to unify these implementations to use IPS specifically for clarity. Longer term, we wish to be able to use the various estimator implementations that have been added.

Since this may be a surprising change to anyone measuring model performance based on the DR estimate we will add this as a deprecation with a flag to force IPS to be used in cases where DR was used before and then in VW 10 we will swap the default to IPS.

It is possible to use the estimators in Python with the vw-estimators library. We need to add documentation about the integration of these packages since it is so important.

[lalo]

we also have these other estimator impls (not part of cb_adf): https://github.com/VowpalWabbit/vowpal_wabbit/tree/master/vowpalwabbit/core/include/vw/core/estimators