Implement scaled dot-product self-attention in NumPy or PyTorch with an optional causal mask.

Project Q, K, and V; reshape to heads; compute QKᵀ/√d; apply causal or padding masks before softmax; multiply by V; then concatenate heads and project. Test tensor shapes and fully masked rows.

Test empty input, one-element input, duplicates, boundary indices, invalid states, and the largest allowed size; state time and space complexity.