Asymmetric Conflict and Synergy in Post-training for LLM-based Multilingual Machine Translation