Improving Time-Series Foundation Model Performance in Data-Sparse Environments Through Cross-Attention Hybrid Fusion