CompactionRL: Reinforcement Learning with Context Compaction for Long-Horizon Agents