A novel approach to evaluating compact finite differences and similar tridiagonal schemes on GPU-accelerated clusters