Kernel digit_last_wdc has race conditions

[fiigii]

[This is a dummy PR to report a bug]

The code section below from kernel digit_last_wdc reads and writes shared memory across different threads without any synchronization.
In the current code, there is no guarantee that the former shared writing in "CALC_LEVEL_SMALL" happens before the later shared reading. You can see this race warning using "cuda-memcheck --tool racecheck"

You may see this program works fine with specific CUDA compiler versions or GPU architectures, but there is no guarantee that works well in the future. So, I suggest adding __syncwarp(); like below. See [1][2] for more details.

if (lane % 16 == 0)
{
	u32 plvl;
	if (lane == 0) plvl = buck[__byte_perm(pair, 0, 0x4510)].hash[1];
	else plvl = buck[__byte_perm(pair, 0, 0x4532)].hash[1];
	slotsmall* bucks = eq->treessmall[1][PACKER::get_bucketid(plvl, RB, SM)];
	u32 slot1 = PACKER::get_slot1(plvl, RB, SM);
	u32 slot0 = PACKER::get_slot0(plvl, slot1, RB, SM);
	levels[lane] = bucks[slot1].hash[2];
	levels[lane + 8] = bucks[slot0].hash[2];
}

__syncwarp();  // suggested change

if (lane % 8 == 0)
	CALC_LEVEL_SMALL(0, lane, lane + 4, 3);

__syncwarp();  // suggested change

if (lane % 4 == 0)
	CALC_LEVEL_SMALL(2, lane, lane + 2, 3);
		
__syncwarp(); // suggested change

if (lane % 2 == 0)
	CALC_LEVEL(0, lane, lane + 1, 4);
		
__syncwarp(); // suggested change

u32 ind[16];

u32 f1 = levels[lane];

[1] https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#synchronization-functions
[2] https://devblogs.nvidia.com/using-cuda-warp-level-primitives/