Dynamic Sparse-Matrix Allocation on GPUs