chore: improve query perf of b-link-tree

  template <typename Warp>

  __device__ static inline bool find(BNode *root, KeyType key, ValueType &result, Warp &warp) {

    auto rank = warp.thread_rank();

    uint64_t rank = warp.thread_rank();

    BNode *next = root;

    while (next) {

      bool isLeaf = next->mLeaf;

      size_t size = next->mSize;

      KeyType nextKey = next->mKeys[rank];

      uint32_t ballot = warp.ballot(nextKey >= key);

      size_t targetIdx = __ffs(ballot) - 1;

      uint16_t isLeaf = next->mLeaf;

      targetIdx = (targetIdx > size) ? size : targetIdx;

      KeyType targetKey = warp.shfl(nextKey, targetIdx);

      uint32_t size = next->mSize;

      uint32_t ballot = warp.ballot(next->mKeys[rank] >= key);

      uint32_t targetIdx = min(__ffs(ballot) - 1, size);

      uint32_t foundKey = next->mKeys[targetIdx] == key;

      if (isLeaf) {

        if (targetIdx < size && targetKey == key) {

        if (foundKey && targetIdx < size) {

          result = next->mValues[targetIdx];

          return true;

        }

      }

      // assumption: the tree is fully constructed and insert / splits are done

      next = next->mChildren[targetIdx + (targetKey == key)];

      next = next->mChildren[targetIdx + foundKey];

    }

    return false;