NVIDIA: optimze v8

- fix that shared memory for fast div is always used even if an algorithm is not using it
- optimize fast div algo
- store `division_result` (64_bit) per thread instead of shuffle around and store it as 32bit