SwiftKV: Fast Prefill-Optimized Inference with Knowledge-Preserving Model Transformation