Improve missing values handing in GBT

[gkronber]

Currently, GBT does not consider missing values (or NaNs in our case) specifically. The training phase uses only comparisons therefore NaN values are implicitly treated as very small (or very large) values (1). When calculating estimated values the output of the right subtree is always used for NaN values.

I think it would be better to do the following

• in training ignore NaN values for input variables
• for estimated values calculation use a weigthed sum when we find a NaN value for a splitting variable (see #2612).

(1) because of the definition that all comparisons to NaN are false.