Mutual Enhancement of Large and Small Language Models with Cross-Silo Knowledge Transfer