Multi-Variable Agents Decomposition for DCOPs to Exploit Multi-Level Parallelism