Evaluating Large Language Models on Multiword Expressions in Multilingual and Code-Switched Contexts