Eider: Empowering Document-level Relation Extraction with Efficient Evidence Extraction and Inference-stage Fusion