allow custom aggregation method for randomForest

[markvanderloo]

by predicting with predict.all=TRUE we get the matrix of predictions. Row-wise aggergation can be customized (but it may cost a lot of memory as the matrix of predictions equals nrow(newdata) x ntree)