fix: cpu_exclusive_scan

[Livinfly]

Firstly, the result after upsweep phase goes wrong. (although not influence the final result)
My experiment result is below.

(two_d < N / 2)
    ./cudaScan -i ones -n 7
    Array size: 7
    1 2 1 4 1 2 1 (upsweep phase)
    2 3 4 6 3 4 3 (downsweep phase)

    ./cudaScan -i ones -n 8
    Array size: 8
    1 2 1 4 1 2 1 4 (upsweep phase)
    0 1 2 3 4 5 6 7 (downsweep phase)

(two_d <= N / 2)
    ./cudaScan -i ones -n 7
    Array size: 7
    1 2 1 4 1 2 1 (upsweep phase)
    2 3 4 6 3 4 3 (downsweep phase)

    ./cudaScan -i ones -n 8
    Array size: 8
    1 2 1 4 1 2 1 8 (upsweep phase)
    0 1 2 3 4 5 6 7 (downsweep phase)

Secondly, as N is not given as power of 2, so during the downsweep, it always visit out-of-bound in output[i + twod1 - 1].

If you want to totally fix it, only choose to transfer N by function like nextPow2 at very first, or just mention at comments, "use it, only when N is power of 2!"