Exposing the Cracks: Vulnerabilities of Retrieval-Augmented LLM-based Machine Translation