G1: update_remset_before_rebuild mixes liveness in words with liveness in bytes

size_t const live_bytes = _cm->liveness(hr->hrm_index());
selected_for_rebuild = tracking_policy->update_before_rebuild(hr, live_bytes);

The code erroneously uses the liveness count (i.e. the amount of live words between bottom and nTAMS) instead of the liveness_in_bytes for the selection of regions for remembered set rebuilding. The effect is that more regions will be selected for rebuild than necessary.

Fix and add a test for G1MixedGCLiveThresholdPercent