Meta Learning of Interface Conditions for Multi-Domain Physics-Informed Neural Networks