Few-Shot End-to-End Object Detection via Constantly Concentrated Encoding Across Heads