The Devil is in the Errors: Leveraging Large Language Models for Fine-grained Machine Translation Evaluation