One-Layer Transformer Provably Learns One-Nearest Neighbor In Context