Hard Negatives, Hard Lessons: Revisiting Training Data Quality for Robust Information Retrieval with LLMs