M-L2O: Towards Generalizable Learning-to-Optimize by Test-Time Fast Self-Adaptation