Currently, when rebuilding remembered sets, one thread takes a whole single region to scan as work unit.

This, particularly for large regions, can defer the pause quite a bit, i.e. increase time to safepoint quite a bit.

Split up the work in parts, each with a yield-point at the end.