Boundary-Guided Learning for Gene Expression Prediction in Spatial Transcriptomics