Learning Problem Decomposition for Efficient Sequential Multi-object Manipulation Planning