UniSH: Unifying Scene and Human Reconstruction in a Feed-Forward Pass