Improve G1UpdateRSOrPushRefClosure

After JDK-8071278 some additional (micro-)optimizations for G1UpdateRSOrPushRefClosure became apparent:

- use the oopDesc::load_heap_oop without decoding for the null check
- one can do the crosses-region check without accessing the heap region
- it's probably better to use the in_cset_state for checking whether something is in the collection set instead of referencing the HeapRegion*
- the same null check and is-in-cset check is repeated in the called G1ParPushHeapRSClosure
- we should not optimize for evacuation failure in this closure, i.e. try to minimize the number of cards to rescan in evacuation failure occurs.