GameEstimator.fit should have the same semantics as spark.ml Predictor fit w.r.t. intercept terms

[fastier-li]

Today, intercept terms in GameEstimator are handled differently than they are in spark.ml Predictor.fit, so that the same DataFrame cannot be passed as is to both to produce the same result. In Game, we need to add an "intercept" feature column if we want to get the same model out of fit as spark.ml. T