Supervised contrastive learning for multi-source domain generalization: Tool failure prediction in data-deficient cryogenic milling processes