Massively Parallel GPU Memory Compaction