Some g1 policy methods for prediction get passed a for_young_only_phase parameter at a fairly high level to reuse it in the various sub-predictions.

Still some callees retrieve the value again from collector state for no apparent reason.

Unify this.