OmniLLP: Enhancing LLM-based Log Level Prediction with Context-Aware Retrieval