T$^3$-S2S: Training-free Triplet Tuning for Sketch to Scene Generation