bugfix: fix cu118 cub usage (#410)

Related issue: sgl-project/sglang#771

This PR fixes the usage of `FlagHeads` cub API in sampling kernels.
As
[documented](https://nvidia.github.io/cccl/cub/api/classcub_1_1BlockDiscontinuity.html),
the default FlagHeads api will always flag the first element, which is
not expected when first element is not `true`.
> For thread0, item input[0] is always flagged.

This PR sets the `tile_predecessor_item` argument (to 0) which will be
compared against input[0].

CUDA 12+ don't have this issue because we are using the new
`SubtractLeft` API instead of `FlagHeads`.

Author: yzh119

include/flashinfer/sampling.cuh

@@ -118,7 +118,7 @@ __device__ __forceinline__ void DeviceSamplingFromProb(
         .SubtractLeft<VEC_SIZE>(greater_than_u, greater_than_u_diff, BoolDiffOp());
 #else
     BlockAdjacentDifference<bool, BLOCK_THREADS>(temp_storage->block_prim.adj_diff)
-        .FlagHeads<VEC_SIZE>(greater_than_u_diff, greater_than_u, BoolDiffOp());
+        .FlagHeads<VEC_SIZE>(greater_than_u_diff, greater_than_u, BoolDiffOp(), 0);
 #endif
     __syncthreads();