Cross-Lingual Stability and Bias in Instruction-Tuned Language Models for Humanitarian NLP