Optimized parallel CUDA scan algorithm to avoid unnecessary writing in the first phase

The original approach (prescan + uniform shift) is more efficient for
inputs that are expensive to evaluate, such as vector expressions.