After remembered set scanning we need to clear the card table as remembered set scanning uses it to decrease overscan during remembered set scan. It and the related temporary data structures are located across several classes and methods. This makes it hard to modify it, and introduces unnecessary code for verification.

Further, the current way of doing this is very slow since it distributes work across threads using a single linked list of elements and a single claim value. In some cases we noticed that the Atomic claim of a new region is more expensive than the actual clearing of the corresponding bitmap.

Move card table clear handling to G1RemSetScanState to hide its implementation from the outside, and tune the thread usage.