pred_contribs=True handling of base_margin

For models with binary:logistic or reg:linear objective functions,
the last column of the matrix obtained by
`model.predict(data, pred_contribs=True) `
, is equal to the mean of the predictions in training data.

But if I `set_base_margin()` on my training data. For a `'count:poisson'` objective, for example, then my last column is equal to
`data.get_base_margin() + 'bias'`

However, the 'bias' is not equal to the mean of predictions in training, i.e.

`bias != np.mean(model.predict(data,output_margin=True) - data.get_base_margin())`

Why is this, and how is base_margin handled together with pred_contribs=True?

The issue names no files or tests. Reproduce the comparison using model.predict(data, pred_contribs=True), model.predict(data, output_margin=True), and data.get_base_margin() for the listed objectives; done means the handling of base_margin and the final contribution column is explained or corrected and covered by an appropriate regression test.

Written by the indexing model from the issue text.