LLM-based Generative Error Correction for Rare Words with Synthetic Data and Phonetic Context